Explore The LLM Database

Mistral Large 4

Mistral AI

Mistral Large 4, also called Le Chonk, is Mistral’s trillion-parameter multimodal flagship, first offered as a public API preview on October 6, 2026. It targets software engineering, long-running agents, visual grounding, and specialist enterprise work including finance, law, and cybersecurity. Training took place on Mistral’s European infrastructure. Its planned weight release is later in October; the preview available at the audit date does not include downloadable weights. Architecture and post-training details beyond the mixture-of-experts configuration remain limited.

MiniCPM-V 4.7 35B-A3B

OpenBMB

MiniCPM-V 4.7 35B-A3B is a public OpenBMB vision-language checkpoint uploaded on October 6. Its official repository contains 16 downloadable BF16 weight shards totaling approximately 35.21B parameters, a hybrid linear/full-attention mixture-of-experts language configuration, and image/video processing metadata. The configuration permits 262,144 positions, but no model card or validated context specification was published in the inspected repository, so no supported context limit is asserted here. Text and visual inputs produce language output through the conditional-generation architecture. The license, training ancestry, evaluated capabilities, and benchmark results remain undisclosed at the audit time; a public weight upload does not supply missing usage rights.

Reka Rho 1

Reka AI

Rho 1 is Reka’s published 19B omni-reasoning research model trained from scratch. A single network unifies language understanding, image and video generation, and robotic actions through shared attention and state. Separate understanding and generation weight streams combine next-token prediction with flow matching. The October 5 release presents architecture and demonstrations as a research preview; it does not establish downloadable weights or a generally accessible API. Its inclusion follows the catalog’s precedent for substantive named published research models.

Kolibri 1

Aleph Alpha

Kolibri 1 is Aleph Alpha’s open-weight German-English reasoning and tool-use model, trained from scratch with a German-focused tokenizer and data mixture. Its hybrid-attention mixture-of-experts Transformer has 78B total parameters and approximately 3.46B active per token. The native 262,144-token context can extrapolate to 1,048,576 tokens, but the publisher recommends remaining within the native window for complex tasks and serving efficiency. Apache-licensed weights support private and sovereign deployments. Sparse activation reduces computation without eliminating the memory needed for the full weights; the published knowledge cutoff is June 18, 2026 for both English and German.

AstaBrief

Allen Institute for AI

AstaBrief is Ai2’s 8B scientific report-generation model, trained from Qwen3-8B with citation-focused supervised fine-tuning and direct preference optimization. It turns a research question and retrieved literature excerpts into a complete cited report in a single pass, and powers Asta’s Fast report mode. Ai2 releases the weights, training data, and a workflow for using local PDFs. The reported latency advantage evaluates the complete report pipeline rather than the checkpoint alone. Citation grounding is improved but does not guarantee that a synthesis preserves every study’s scope or conclusions.

Solar Mini 4

Upstage

Solar Mini 4 is Upstage’s compact agent model for high-volume retrieval, structured generation, and tool-driven workflows. Its mixture-of-experts architecture has 35B total parameters and 3B active per token, with a 512K context and up to 128K output tokens. It supports parallel function calling and is offered through Upstage’s hosted API, with on-premises deployment options. The model emphasizes compute efficiency for repetitive work while retaining long-context reasoning. On-premises availability does not imply publicly downloadable weights; memory requirements still reflect the full 35B parameters.

Apex Flash 1

Cantina Security

Apex Flash 1 is Cantina Security and Yeta’s reinforcement-learning post-training of GLM-5.3 Flash for focused software security investigations. It reads code, emits tool calls, investigates vulnerabilities, and verifies results in running targets. MIT-licensed weights are available, retaining the base model’s image-text architecture, although published evaluations cover text-based security tasks only. It is intended as a focused worker inside an agent harness. The standard checkpoint and an experimental abliterated derivative are distinguished; reported evaluation results apply to the standard checkpoint.

Ling 3.1 Flash

InclusionAI (Ant Group)

Ling 3.1 Flash is InclusionAI’s hosted hybrid reasoning language model for coding, multi-step analysis, and agents that call tools. Its mixture-of-experts network has 560B total parameters and activates 25B per token. The verified September 30 availability announcement provides a 262K context through Vercel AI Gateway, suitable for long documents, repository analysis, and extended task histories. Standard and temporary free API identifiers represent the same model. Public weights were not verified at the audit date, so hosted availability is recorded without assuming a downloadable license or checkpoint.

Mercury Voice

Inception

Mercury Voice is a diffusion language model tuned for the language-reasoning component of real-time voice agents. It generates text, reasons, and calls tools; speech recognition and synthesis are supplied by the surrounding voice stack. Three reasoning-effort settings balance response quality and latency. Enterprise API access became generally available after an earlier preview announcement. Public weights and parameter counts are undisclosed. Reported latency depends on hardware, concurrency, and prompt distribution; production applications require workload-specific evaluation.

GPT-6.1 Sol

OpenAI

GPT-6.1 Sol is an independently released update to the GPT-6 Sol tier, targeting capability near Astra at a lower price. It supports complex coding, computer use, and professional work with text and image inputs. Reasoning is always enabled at one of low, medium, high, xhigh, or max effort; tool calling requires Responses. The documented context and output limits remain those of the GPT-6 family.

Claude Sonnet 5.5

Anthropic

Claude Sonnet 5.5 is the second released Claude 5.5 tier, focused on faster everyday agent work, bug fixes, visual design, and polished documents. It improves over Sonnet 5 while reducing typical token consumption. A million-token context and adaptive thinking support sustained tasks. The release changes thinking configuration, forced tool use, and computer-tool compatibility, so integrations need model-specific migration checks.

Holo4 27B

H Company

Holo4 27B is H Company’s computer-use vision-language model built on Qwen3.8-27B. It is post-trained to work across graphical interfaces, code, MCP tools, and APIs, using screenshots and tool results to choose its next actions. Released weights support local agent research; an external harness executes clicks, typing, and code and returns observations. The dense 27B model uses a non-commercial weight license, unlike the Apache-licensed sparse Holo4 tier. Its card specifies a 262,144-token configuration limit; this is not a promise of consistent task performance throughout that window.

Holo4 35B-A3B

H Company

Holo4 35B-A3B is H Company’s computer-use vision-language model built on Qwen3.6-35B-A3B. It is post-trained to work across graphical interfaces, code, MCP tools, and APIs, using screenshots and tool results to choose its next actions. Released weights support local agent research; an external harness executes clicks, typing, and code and returns observations. The sparse model activates about 3B language parameters per token but still needs memory for its full checkpoint. Its card specifies a 262,144-token context. Apache-licensed weights and hosted H Models API access are available.

Holotron4 Nano

H Company

Holotron4 Nano is H Company’s computer-use vision-language model built on NVIDIA Nemotron 3 Nano Omni. It is post-trained to work across graphical interfaces, code, MCP tools, and APIs, using screenshots and tool results to choose its next actions. Released weights support local agent research; an external harness executes clicks, typing, and code and returns observations. The sparse model activates about 3B language parameters per token but still needs memory for its full checkpoint. Its card specifies a 262,144-token context. The NVIDIA Open Model Agreement governs its weights; the released page establishes text and image understanding without assuming every modality of its Omni backbone.

Naive-N0.5-Flash

NaiveAI

Naive-N0.5-Flash is NaiveAI’s open-weight coding and AI research model with 309B total parameters and 15.5B active per token. Continued pre-training adapts MiMo-V2.5 by replacing global attention with a hybrid of sliding-window attention and lightweight DeepSeek Sparse Attention. No full-attention layers remain, and the model supports a native 1M context. MIT-licensed FP8 weights and deployment code are available. Its research focus includes experiment execution, kernel optimization, and software development through external tools. The launch advertises future API access; invitation or future access is not treated as an unrestricted production API.

IndustryLLM

Alibaba

IndustryLLM is Alibaba’s open-weight language model for industrial procurement and engineering-standard reasoning. It adapts Qwen3.5-35B-A3B-Base through continued pre-training on industrial documents and transactions followed by task-focused supervised fine-tuning. The 35.95B-parameter hybrid mixture-of-experts network activates approximately 3B per token. Training targets informal buyer language, ambiguous product codes, contradictory specifications, and evidence-based constraint evaluation. Apache-licensed BF16 weights are published. Its 262K architectural context is distinct from 12,288-token packed fine-tuning sequences; this record uses the supported training length conservatively. The September 25 paper predates later discovery-index coverage.

Ember-1

Fireworks AI

Ember-1 is Fireworks AI’s specialized Kimi K3-derived reasoning model, offered through its serverless API and playground. The 2.78T-parameter mixture-of-experts model supports image inputs, function calling, and a context advertised as 1040K tokens. It is optimized for shorter reasoning traces: the publisher reports roughly 40% fewer generated tokens with comparable results on its own evaluations. This is a provider evaluation claim rather than a guarantee of latency or quality in every agent workload. Ember-1’s September 23 launch establishes public availability, while the catalog’s September 22 creation timestamp is earlier metadata. Public model weights are not offered.

Claude Opus 5.5

Anthropic

Claude Opus 5.5 is the first released Claude 5.5 tier, designed for sustained agentic coding and professional work. Anthropic positions it near Fable 5.1 on most workloads with substantially lower cost than Opus 5, partly through token efficiency. It has always-on adaptive thinking, a million-token context, text and image input, and broad tool access. Effort settings control the latency and computation devoted to a task.

GPT-6 Luna

OpenAI

GPT-6 Luna is the efficient GPT-6 tier for focused high-volume tasks. It supports text and image input, text output, and reasoning effort from none to max. Responses exposes its built-in tools and function calling; Chat Completions function calling requires none effort. The October Decisions API adds a typed-output interface using this same model identity.

GPT-6 Sol

OpenAI

GPT-6 Sol is the GPT-6 workhorse tier for complex coding and agent workflows, released alongside Luna. It combines a long context with image understanding and reasoning effort from none to max. Responses provides its built-in tools, while Chat Completions function calling requires reasoning to be disabled. GPT-6.1 Sol is a subsequent distinct weight update.

MiMo V2.6 Flash

Xiaomi

MiMo V2.6 Flash is Xiaomi’s lower-compute multimodal agent tier. Its sparse model has 309B total parameters and 15B active per token, with native text, image, audio, and video understanding and a 1M context window. Reinforcement learning scales task environments and feedback across coding, general work, visual tasks, and security research. MIT-licensed weights and hosted inference support private deployment and API applications. Subsequent MOPD checkpoints are documented variants of the same named tier.

MiMo V2.6 Pro

Xiaomi

MiMo V2.6 Pro extends Xiaomi’s flagship into native multimodal agent reasoning over text, images, audio, and video. It retains a 1.02T sparse backbone with 42B active parameters and a 1M-token context. Large-scale reinforcement learning uses verifiable coding, knowledge-work, visual, and security environments. API access and MIT-licensed RL checkpoints were released together. Later MOPD refinements and UltraSpeed serving are grouped as variants; benchmark results depend on the chosen checkpoint and agent harness.

Altar 1

Aikido Security

Altar 1 is Aikido Security’s independently named, compressed language model for private defensive security infrastructure. It prunes GLM-5.3 to 168 routed experts per layer and stores most expert weights in INT4, reducing the downloadable checkpoint to about 328 GB and 504B parameters without adding new trained skills. Calibration emphasizes security, coding, tool use, and multiple languages. The primary model card documents a 128K serving setup on four H200 GPUs. It remains governed by the GLM-5.3 license and requires external tools for vulnerability investigation.

Grok 4.7

xAI

Grok 4.7 is a proprietary xAI language-model release. The September frontier update improves coding, agentic tasks, and knowledge work and is available through the API. It uses text and image inputs and text output, configurable reasoning effort, and tool interfaces. A faster serving option uses the same underlying model and is recorded as a deployment variant, not a separate checkpoint. Weights are not publicly downloadable. Published agent and coding results depend on the evaluation harness, tool permissions, and inference settings, so no unsupported standalone benchmark score is assigned.

Qwen-Audio-3.1-Realtime

Alibaba

Qwen-Audio-3.1-Realtime is a proprietary language-capable speech conversation model for fluid real-time dialogue. It combines speech reasoning with duplex interaction, allowing conversational responses rather than only transcription or speech synthesis. The 3.1 Plus update supports a 262,144-token context, function calling, web search, voice cloning, and additional system voices, while retaining the 3.0 Plus integration protocol. Hosted streaming APIs provide access; no downloadable weights or checkpoint parameter count is verified.

Qwen3.8-Omni-Flash

Alibaba

Qwen3.8-Omni-Flash is a native omnimodal agent model built on the Qwen3.8-Flash-Next architecture. It accepts text, images, audio, and video in a million-token context, extending coding, GUI operation, and knowledge work to audio-visual productivity. Standard APIs return text, while the subsequently available Realtime variant supports text and speech output, multichannel audio, video aggregation, and remote MCP tools. Hosted access and companion plugins support video editing, multimedia summarization, and real-time conversations; its weights are not publicly released.

Qwen3.8-LiveTranslate-Flash

Alibaba

Qwen3.8-LiveTranslate-Flash is a language-capable real-time interpretation model that uses speech and visual context to translate conversations. It recognizes 60 source languages and supports spoken output in 29 languages, with text and audio responses exposed through the Realtime API. Visual grounding helps resolve terminology and references during meetings, classes, and livestreams. This update improves latency and enables real-time speaker diarization and transcription by default; the September 17 API changelog precedes the following launch article. Weights are not publicly downloadable.

Bonsai 2 27B

PrismML

Bonsai 2 27B is PrismML’s second-generation vision-language flagship, derived from Qwen3.8-27B. Ternary language weights and a hybrid linear/full-attention backbone reduce deployment footprint while retaining language reasoning, coding, image understanding, and agent-oriented capabilities. The current card specifies 27.36B total parameters, 262K context, and two GGUF packing formats; the vision encoder ships separately. Apache-licensed weights became available September 17. The September 24 smart-glasses coverage concerns deployment of Bonsai technology rather than a new named checkpoint. Compression quality and throughput claims depend on evaluation setup, hardware, packing, and runtime support.

Gemini 3.8 Live Extended Thinking

Google

Gemini 3.8 Live Extended Thinking is the separately named high-reasoning audio-to-audio tier for live interactions. It spends more computation on background reasoning while handling text, image, audio, and video input and generating text and audio. Asynchronous function calls allow tool-driven work during a session. Integrations must monitor interaction status because completion of a response turn does not necessarily end background processing.

Gemini 3.8 Live

Google

Gemini 3.8 Live is a generally available real-time audio-to-audio model for low-latency voice agents and dialogue. Text, image, audio, and video input can be combined in a live session, with text and audio responses. Interleaved reasoning and asynchronous function calling help agents continue an interaction while using tools. It is the standard Live tier; extended background reasoning has a separately named model identity.

Koa

Salesforce

Koa is Salesforce’s enterprise language model for CRM reasoning and multi-turn tool use, built by post-training NVIDIA Nemotron 3 Super 120B. Supervised fine-tuning and GRPO reinforcement learning use public and synthetic workflow data rather than Salesforce customer records. Persona-conditioned tasks and tool-grounded rewards teach the model to follow business procedures in sales, service, and related enterprise workflows. Salesforce hosts inference within its infrastructure and offers access to selected Agentforce pilot customers at release; general availability was still expected for winter 2026. Public weights and a standalone model API are not established.

Atria Dawn Preview

Shanghai AI Laboratory

Atria Dawn Preview is Shanghai AI Laboratory’s agentic model for turning research and engineering questions into executable results. Built on GLM-5.2’s 744B mixture-of-experts foundation, it is post-trained for discovery, code creation, office deliverables, and cybersecurity tasks in authorized environments. The released text-only checkpoint supports a documented 256K inference context, downloadable MIT-licensed weights, and hosted API access. It combines tool execution with environmental feedback for experiment design, failure recovery, and verification. Configuration maximums above the documented serving limit are not treated as evidence of a larger supported context.

Intern S2 397B

Shanghai AI Laboratory

Intern S2 is the stable scientific multimodal follow-up to the 397B preview checkpoint. It strengthens extended agent execution and research workflows, with deep-thinking and ordinary-response modes. Public API documentation identifies the stable model separately and schedules preview decommissioning for October 31. The release includes downloadable Apache-licensed weights and chat access. The official laboratory announcement identifies the public release date as 2026-09-13.

Agnes 3.0 Flash Preview

Agnes AI

Agnes 3.0 Flash Preview is a 33B open-weight multimodal language model, originally published without the Preview suffix. Its hybrid decoder alternates three gated delta-rule recurrent layers with one global attention layer, reducing the number of layers with growing key-value caches. The Apache-licensed checkpoint accepts text, images, and video and generates text with 262,144-token context, configurable reasoning effort, and tool-call formatting. The provider explicitly identifies these weights as an earlier checkpoint distinct from the production Agnes 3.0 Flash API model; hosted model benchmarks and larger context claims do not apply to this Preview.

Kimi K2.8 Preview

Moonshot AI

Kimi K2.8 Preview is Moonshot AI’s public coding-agent preview, rolled out to Kimi Code users on September 11. It improves coding and agent execution and reduces overthinking relative to K2.7 Code, with low, high, and max reasoning effort and a million-token context. Official model configuration confirms text, image, and video input. The existing kimi-for-coding alias now selects this distinct model rather than K2.7 Code. No public weights, parameter count, or detailed architecture are established; the membership API and supported coding clients provide access.

SWE-2

Cognition

SWE-2 is Cognition’s Kimi K3-derived coding model, first publicly introduced and served in Devin Desktop and CLI on September 10, 2026. It extends Cognition’s reinforcement-learning recipe to the roughly 2.8T-parameter scale, optimizing several reasoning-effort levels in one training run. Effort-specific cost penalties target the tradeoff between engineering capability and inference cost, while expanded coding environments and hardened verifiers encourage more reliable task completion. The released identity is the trained model rather than the surrounding Devin application. Web and Fusion rollout were announced separately. Public weights and exact model-serving context/output limits are not established by the launch announcement.

DeepSeek-V4.1-Flash

DeepSeek

DeepSeek-V4.1-Flash is a natively multimodal MoE model trained from scratch on 45 trillion tokens. It uses a causal encoder-decoder layout, Compressed Sparse Attention 2, Engram conditional memory, and DSpark speculative decoding to reduce inference and KV-cache cost. Its backbone activates different parameter counts during prefill and decoding, while a separate 196B conditional-memory table is sparsely accessed. The model accepts images and text and generates text, supports million-token contexts and adjustable reasoning effort, and powers the official deepseek-flash API. MIT weights and model-specific inference code support self-hosting.

GPT-Live 1

OpenAI

GPT-Live 1 is a language-capable full-duplex voice model that can listen and speak simultaneously for fluid conversation and interruption handling. Complex reasoning and tools can be delegated to a backend model or agent rather than blocking the voice interaction. Its dedicated Live API uses session-duration pricing, with backend computation billed separately. It accepts audio and text, without image or video input.

North Small Translate

Cohere

North Small Translate is a dedicated text translation model supporting more than 50 languages. Its 218B-parameter mixture-of-experts network activates 25B parameters and targets long-form translation with efficient serving. Open weights are available in several precisions under CC BY-NC 4.0, while commercial use requires appropriate licensing. It can be evaluated through Cohere’s Chat API and a public demo. The release date follows the September 9 API changelog; the launch blog followed on September 10. Translation quality claims depend on the specific language pair, domain, and evaluation method.

Tiny Aya Thinker

Cohere

Tiny Aya Thinker is Cohere Labs’ 3.35B multilingual reasoning research release, published with the Building Multilingual Bridges paper. L2-Thinker produces reasoning in the user’s language across 60 languages, while En-Thinker provides an English-reasoning comparison. The models build on a 32K Tiny Aya base checkpoint. Downloadable research weights and a public demonstration support multilingual reasoning experiments. The paper publication precedes the October 5 explanatory blog; that later blog is not the first release.

LLaDA-UI

InclusionAI (Ant Group)

LLaDA-UI brings block-wise diffusion generation to a vision-language GUI model. The approximately 16.7B MoE model preserves screenshot aspect ratios and visual detail, reasons about natural-language tasks, and produces grounded coordinates or tagged reasoning and structured actions for mobile, desktop, and web interfaces. The September 9 technical report precedes the September 15 weight publication, independently of the repository creation timestamp. Public weights and serving examples are available; the exact weight license is not declared in the current card and is not inferred from neighboring LLaDA models.

Mercury 2.5

Inception

Mercury 2.5 is Inception’s updated hosted diffusion reasoning model for coding agents, structured generation, and low-latency conversational workflows. It improves the quality and serving tradeoff over Mercury 2 while preserving parallel refinement. API and chat access are publicly available, with enterprise deployment options. Public weights and parameter counts are undisclosed. Reported latency depends on hardware, concurrency, and prompt distribution; production applications require workload-specific evaluation.

Gander

Tencent

Gander is a 9B full-duplex audio-visual interaction model developed by Tencent’s Hunyuan Speech team with academic collaborators. Its MiniCPM-o 4.5-derived Thinker and streaming Talker handle speech, images, video frames, text, turn taking, interruption, and structured delegation while tasks continue asynchronously. Apache-licensed model components and a serving runtime are publicly available. The larger system pairs this conversational model with a separately configured reasoning agent for long-horizon execution; those back-end model capabilities are not assumed to reside inside the 9B checkpoint. The release date follows the September 8 technical report that publishes the named model; the project dates its checkpoint release to September 9.

Occamy 1.0

Alibaba

Occamy 1.0 is Alibaba Accio’s open-weight co-work model, first described in its September 4 research paper. Further training of Qwen3.6-35B-A3B targets long, stateful workflows using search, code, files, structured APIs, and productivity applications. Its 35B mixture-of-experts backbone activates approximately 3B parameters and retains image input. Full-parameter supervised training, preference optimization, model merging, and sustained-action optimization emphasize recovery and follow-through. Apache-licensed BF16 weights and routine deployment formats are available. The architecture supports 262K context, while supervised training uses 131K sequences; this record uses the latter conservative operating scale.

Ling-3.0-Flash-VL

InclusionAI (Ant Group)

Ling-3.0-Flash-VL extends the Ling3.0-Flash language backbone with native image and video understanding for visual reasoning and agent execution. A ViT encoder, MLP projector, VideoRoPE, and hybrid KDA/gated MLA backbone combine visual evidence with planning, action, and verification. The 124B model activates 5.5B parameters and supports up to 256K context through the documented deployment extension. The September 4 official changelog establishes public chat/API access before the subsequent open-weight promotion. MIT weights enable local visual assistants and document or video analysis.

Iris-mini

AllSpark Research

Iris-mini is AllSpark Research’s open-weight search-specialized language model, published in the September 3 Iris paper. It is post-trained from Qwen3.6-35B-A3B through alternating supervised fine-tuning and reinforcement learning, learning what to search, how to read evidence, and when enough information has been collected. Its mixture-of-experts backbone has 35B total and 3B active parameters, with a 256K context. Apache-licensed weights and an evaluation harness are available. The checkpoint generates reasoning, tool calls, and final language answers; search itself is performed by external tools. Published scores depend strongly on context resets and retry policies.

Iris-pro

AllSpark Research

Iris-pro is AllSpark Research’s open-weight search-specialized language model, published in the September 3 Iris paper. It is post-trained from Qwen3.5-397B-A17B through alternating supervised fine-tuning and reinforcement learning, learning what to search, how to read evidence, and when enough information has been collected. Its mixture-of-experts backbone has 397B total and 17B active parameters, with a 256K context. Apache-licensed weights and an evaluation harness are available. The checkpoint generates reasoning, tool calls, and final language answers; search itself is performed by external tools. Published scores depend strongly on context resets and retry policies.

K2 Horizon Uno

Institute of Foundation Models

K2 Horizon Uno augments the Horizon autoregressive language backbone with a conditional LoRA diffusion pathway for parallel token proposals. The released 7B and 0.9B adapters require separately downloaded base model weights; the adapter repository is not a complete 7B language checkpoint. Uno’s sampler verifies proposed tokens against the frozen autoregressive pathway, preserving its output distribution under the method’s assumptions. This is a substantive decoding specialization rather than an independently enlarged language backbone. Apache-licensed adapters, code, training recipes, and a technical report support evaluation; throughput gains depend on workload and inference configuration.

K2 Horizon

Institute of Foundation Models

K2 Horizon is the Institute of Foundation Models’ open language-model fleet spanning 0.9B, 3.7B, 7B, 32B, 36B-A4B, and 375B-A23B tiers. The release combines dense and sparse models, including the 36B tier’s Mixture-of-Value-Attention design, for reasoning, mathematics, coding, and agents. Apache-licensed models and code are accompanied by intermediate checkpoints, training data or construction recipes, configurations, logs, and evaluations. Context differs by training stage: the documented 7B and 32B final views support 512K, while earlier base and midtraining checkpoints have smaller limits. Tier names describe core model sizes and may differ from all-parameter repository totals.

Nemotron Labs 3 Competitive Coding

NVIDIA

Nemotron Labs 3 Competitive Coding is an open-weight NVIDIA language-model release. A competitive-programming specialist trained from the Nemotron 3 Ultra base model with supervised learning and reasoning reinforcement learning. It generates candidate algorithmic solutions and supports research on difficult coding contests. The public checkpoint is NVFP4. Contest results involving multiple candidate searches or GenCorrect evaluate a system rather than a single response. The license permits commercial and non-commercial use subject to its published terms.

Nemotron Labs Ultra Math

NVIDIA

Nemotron Labs Ultra Math is an open-weight NVIDIA language-model release. NVIDIA’s released olympiad mathematics specialists build on Nemotron 3 Ultra for rigorous proof generation, checking, and self-evaluation. The SFT and reinforcement-learning checkpoints are grouped here. They were used in an ensemble with inference-time scaling for IMO-level problem solving; ensemble medal-level claims are not standalone model accuracy. The license permits commercial and non-commercial use subject to its published terms.

GPT-6 Astra

OpenAI

GPT-6 Astra is OpenAI’s frontier GPT-6 model for difficult end-to-end work, combining reasoning, coding, computer use, research, and document creation. It accepts text and images and supports effort levels from low through max. Tool calls require the Responses API; Chat Completions serves text generation without tools. It includes support for extended workflows and mid-turn steering through the API.

Gemini 3.8 Flash

Google

Gemini 3.8 Flash is the stable September update to the Flash line, improving sustained coding, instruction following, and reliable multi-step agent work. It retains multimodal input and a million-token context with text output. Native tools cover search grounding, code execution, function calling, and preview computer use. Low, medium, and high thinking levels balance response time against reasoning computation.

Muse Spark 1.3

Meta

Muse Spark 1.3 is a proprietary Meta multimodal reasoning model for coding and agentic knowledge work. The 1.3 update improves long instruction adherence, multi-task coordination, self-awareness of task limits, and cooperation with users during extended work. It accepts text, images, video, and audio and produces text; the API specifies a 1,048,576-token context for these versions. Standard and, where offered, Contributor access use the same model capabilities but different data-use terms. Contributor access is a serving tier rather than a separate model. Future open-weight plans are not a weight release.

Claude Fable 5.1

Anthropic

Claude Fable 5.1 is the independent update to Fable 5 for long-running coding, knowledge work, and research. It improves grounding, root-cause debugging, and efficiency at lower effort. Claude Mythos 5.1 is the same model with more permissive safeguards for verified cybersecurity and life sciences work, represented as a variant. Adaptive thinking is always enabled, and a million-token context supports extended tool-assisted tasks.

Agnes 3.0 Flash

Agnes AI

Agnes 3.0 Flash is Agnes AI’s hosted production model for coding agents, reliable tool orchestration, long-running task execution, and instruction following. Its API supports text and image URL input with text output through Chat Completions, Responses, and Messages-compatible interfaces. The current API guide specifies 512K context and a 65,536-token output limit. The provider’s separate Preview weight card describes a 1M production configuration, so this page follows the lower documented serving contract. Production parameters and architecture are undisclosed, and the downloadable 33B Apache Preview is explicitly a different checkpoint. September month precision avoids assuming that repository publication dates describe the hosted launch.

HuatuoGPT-3

FreedomIntelligence

HuatuoGPT-3 is FreedomIntelligence’s named medical language-model series, comprising 9B and 27B checkpoints based on Qwen3.5-9B and Qwen3.8-27B respectively. One-stage Policy Optimization adapts the models through medical reinforcement learning without preceding domain-specific supervised fine-tuning, using teacher guidance that is retired as performance improves. Apache-licensed weights, training code, data, and a rubric grader are published. Inference requires thinking mode and generates reasoning before a final response. Vision-capable backbones are retained, but the released model-card examples and medical evaluations focus on textual questions. September 2026 weight publication is verified; first-day precision remains unresolved.

IQuest-Q1

IQuestLab

IQuest-Q1 is an open-weight foundation language model for agentic coding, reasoning, and multi-step tool use. It combines 320B total parameters with approximately 15B activated per token, interleaving sliding-window and full attention across an 88-layer mixture-of-experts backbone. The documented context limit is 512K; client settings that label the model as 1M do not extend that limit. Its English and Chinese model card includes local deployment examples and an IQuest-specific tool-call format. The modified MIT license permits commercial use but requires prominent IQuest-Q1 attribution in commercial product interfaces. September 2026 publication is verified; the first public launch day is not independently established.

Nex-N2.5 Max

Nex AGI

Nex-N2.5 Max is a separately positioned tier in Nex AGI’s September 2026 agentic family. The Max tier is Nex AGI’s first trillion-parameter-scale post-training effort. Unlike mini and Pro, its 1.6T-parameter foundation is text-only, emphasizing coding, tool execution, scientific research, and complex productivity tasks. Adaptive thinking can produce direct answers or reasoning traces, with controllable reasoning effort. Apache-licensed downloadable weights and compatible serving examples are published. The official repository establishes the released family but does not date its first public availability precisely, so month precision is preserved. Agent evaluation results reflect the model together with tools, context management, and execution budgets.

Nex-N2.5 mini

Nex AGI

Nex-N2.5 mini is a separately positioned tier in Nex AGI’s September 2026 agentic family. The mini tier preserves multimodal understanding while focusing on responsive computer use, web browsing, and visually grounded actions. It provides a smaller weight footprint than Pro and Max. Adaptive thinking can produce direct answers or reasoning traces, with controllable reasoning effort. Apache-licensed downloadable weights and compatible serving examples are published. The official repository establishes the released family but does not date its first public availability precisely, so month precision is preserved. Agent evaluation results reflect the model together with tools, context management, and execution budgets.

Nex-N2.5 Pro

Nex AGI

Nex-N2.5 Pro is a separately positioned tier in Nex AGI’s September 2026 agentic family. The Pro tier strengthens multimodal long-horizon tasks, including computer use, web browsing, coding, and knowledge work. Its large mixture-of-experts checkpoint targets greater capability at higher serving cost. Adaptive thinking can produce direct answers or reasoning traces, with controllable reasoning effort. Apache-licensed downloadable weights and compatible serving examples are published. The official repository establishes the released family but does not date its first public availability precisely, so month precision is preserved. Agent evaluation results reflect the model together with tools, context management, and execution budgets.

Intern Lumina U2

Shanghai AI Laboratory

Intern Lumina U2 is a unified language-capable multimodal diffusion model covering text questions, visual understanding, image generation, and editing. Its sparse 16B backbone activates 1B parameters and uses an eight-codebook discrete visual tokenizer. The Ascend-trained weights are publicly available; the NVIDIA-trained checkpoint, training code, and report remain future work. Code publication in August preceded the September weight release. Primary catalogs verify the release month; the exact first-public day remains unverified and is not inferred from private repository or API creation metadata.

Step 5 Preview

StepFun

Step 5 Preview is StepFun’s hosted flagship for agentic coding, research, professional knowledge work, and financial analysis. Its sparse mixture-of-experts architecture has 600B total parameters and 27B active per token. The documented API accepts text, images, and video, generates text, and supports a 1M-token context with up to 64K output. Tool calls, JSON Schema output, caching, and low, medium, or high reasoning effort support sustained workflows in an external application. It was publicly announced on September 20, 2026; September is used because earlier reported API availability dates are not resolved by a first-party dated changelog. Announced October 15 weights are future availability at this audit cutoff.

Qwen-Drive 1.0

Alibaba

Qwen-Drive1.0 is a vision-language foundation model specialized for autonomous-driving research while retaining free-form visual question answering and instruction following. Its Qwen3.5-4B language/vision backbone is paired with an external bird’s-eye-view perception head and a flow-matching Planning Expert for future trajectories. General vision-language supervision is mixed with driving data to preserve broad visual understanding. The August31 technical report precedes the public weight rollout. Apache-2.0 weights and demo data support research; SFT and reward-optimized planners are grouped as variants.

Qwen-Flash-Character

Alibaba

Qwen-Flash-Character is a hosted language-model specialization for multilingual role-playing and anthropomorphic conversation. The official model-release catalog describes improved character consistency, context-aware dialogue progression, and empathetic engagement, supporting personalized fictional personas and sustained narrative interactions. This named model identity is distinct from a role-play application or general Flash alias. Its weights, parameter count, and detailed architecture are not publicly established; access follows the provider’s hosted service documentation.

Hy4 Preview

Tencent

Hy4 Preview is Tencent’s open-weight productivity-focused language model with 770B total parameters, 49B active per token, and a 1M-token context window. It targets code planning, debugging, office workflows, financial analysis, and research tasks that require sustained tool interactions. Tencent released weights under Apache 2.0 and made the model available through its productivity applications and inference APIs. Its preview status distinguishes this release from future Hy4 checkpoints; planned follow-up models are not assumed available. Reported improvements in internal task evaluations are provider findings rather than independent measurements.

UI-Venus-2

InclusionAI (Ant Group)

UI-Venus-2 is a distinct vision-language model specialization for GUI interaction, combining screenshot understanding with natural-language task reasoning and structured actions. The second generation adds multilingual mobile, browser, and desktop environments, keypoint-grounded verification, and reflection-based recovery. The released 9B checkpoint is initialized from Qwen3.5-9B; a 27B tier is described in research evaluations, but its weights are not verified. The current model card explicitly leaves its weight license unresolved. Local inference requires an external action parser and execution framework; the model itself emits language reasoning and actions.

Qwen3.8-Flash-Next

Alibaba

Qwen3.8-Flash-Next introduces an efficiency-oriented architecture for multimodal reasoning and agents. It combines Gated DeltaNet with Qwen Sparse Attention, plus 51B n-gram embedding parameters held outside the accelerator and 4B multi-token-prediction parameters. The 125B backbone activates 6B parameters per token. Public weights use the Qwen Community License 1.0 and support a native 256K context with optional million-token extension. The hosted Qwen3.8-Flash service adds production features and has a separate page.

Qwen3.8-Flash

Alibaba

Qwen3.8-Flash is the hosted production service based on Qwen3.8-Flash-Next. It adds a million-token context by default and official built-in tools to the efficiency-focused multimodal architecture. Text, images, and video support code assistance, GUI agents, and long-document understanding, with OpenAI- and Anthropic-compatible APIs. The downloadable Flash-Next checkpoint is separately cataloged because its native context and weight-license profile differ from hosted service access.

GLM-5.3-Flash

Z.ai

GLM-5.3-Flash is a natively multimodal GLM model built on a newly trained base. Its hybrid sparse and linear attention, manifold-constrained hyper-connections, and 30T-token multimodal pretraining target strong capability at lower long-context serving cost. Native visual perception supports agents that observe interfaces, inspect rendering results, and refine code. The 320B model activates 18B parameters per token and supports million-token contexts. MIT weights and official API/chat access provide both local and hosted options.

Granite 4.2

IBM

Granite 4.2 adds native reasoning and reasoning-guided tool use to IBM’s open enterprise language models. The release includes 3B, 8B, and 30B checkpoints with full, low-effort, and non-thinking modes, allowing developers to select a latency and reasoning tradeoff. Multi-stage reinforcement learning targets coding, mathematics, science, and agent execution. All sizes natively support 128K context; the 30B checkpoint additionally supports a documented extension to 512K. Apache-licensed weights enable private deployment and customization. The family page groups sizes while preserving their individual checkpoint names as variants.

DeepSeek-V4-Flash-Vision-Exp

DeepSeek

DeepSeek-V4-Flash-Vision-Exp is the experimental vision-language specialization of the production V4 Flash model. It adds image understanding for agent tasks while retaining the text capabilities of V4 Flash, enabling workflows that inspect interfaces, diagrams, and other visual context. It was publicly accessible through a dedicated DeepSeek API model name before downloadable MIT weights were released. Following the September 10 V4.1 Flash launch, the experimental endpoint was retired and its alias redirected; the original checkpoint remains available for self-hosted research.

Qwen3.8 27B

Alibaba

Qwen3.8-27B is the dense open-weight tier of the Qwen3.8 release, with native vision support and a hybrid-attention language backbone. It targets coding, professional work, reasoning, and visual agents at a substantially smaller deployment footprint than the Max-scale checkpoint. Reasoning effort and preserved thinking support multi-turn agent workflows. Apache-2.0 weights allow self-hosting with a native 256K context and optional extension. Its multimodal capabilities and license are specific to this dense tier, differing from the text-only 2.4T open checkpoint.

Nemotron Labs Teachers

NVIDIA

Nemotron Labs Teachers is an open-weight NVIDIA language-model release. A released family of domain-specialized Nemotron 3 Ultra teacher checkpoints. Chat, instruction following, general reasoning, STEM, and competitive coding teachers provide differentiated training signals for multi-teacher on-policy distillation. The chat checkpoint is also a standalone reasoning and conversation model, not merely a reward scorer. The license permits commercial and non-commercial use subject to its published terms.

MathForm-8B

OpenBMB

MathForm-8B is OpenBMB’s language-to-code model for translating natural-language mathematical statements into Lean 4. It fine-tunes Qwen3-8B on FormalVerse with supervised learning followed by reinforcement learning using compilation and semantic-consistency feedback. The research framework retrieves Mathlib definitions and revises generated formalizations using diagnostics, supporting more faithful mathematical translation. Apache-licensed weights, approximately 367K verified examples, and evaluation code accompany the August 14 paper. This is a named mathematical language specialization; the model is distinct from the surrounding retrieval and verification pipeline. Successful compilation alone does not establish that a formal statement preserves the intended proposition.

GLM-5.3

Z.ai

GLM-5.3 advances coding, long-horizon engineering, and vulnerability-analysis capabilities through post-training on the GLM-5.2 base model. The model remains text-based and supports a million-token context, selectable reasoning effort, and persistent thinking history. Z.ai initially announced hosted availability before releasing downloadable weights. Those weights use the custom GLM-5.3 License rather than the MIT license of earlier GLM-5 models. It is distinct from GLM-5.3-Flash, which introduces a new native multimodal architecture.

Gemini 3.7 Flash

Google

Gemini 3.7 Flash is a stable model update with substantial improvements in software engineering, web development, and tool-based workflows. It accepts text, images, audio, and video and generates text, with a million-token input context. Low, medium, and high thinking levels control reasoning computation; native function calls, grounded retrieval, and preview computer use support agent integrations.

MAI-Cyber-1-Flash

Microsoft

MAI-Cyber-1-Flash is a compact security model in Microsoft’s MAI-Thinking-1 lineage, offered to verified defenders through the MDASH vulnerability-analysis and remediation harness. It handles routine code-security tasks while MDASH routes difficult cases to larger models. The documented August release is a controlled enterprise offering rather than a general downloadable or standalone public API model. Reported CyberGym results evaluate a combined system using this model and GPT-5.4, so those numbers are not assigned to this model alone.

Qwen3.8 2.4T-A95B

Alibaba

Qwen3.8-2.4T-A95B is the open-weight Max-scale Qwen3.8 checkpoint, with 2.4T total parameters and 95B activated per token. It advances coding, professional work, research, and long-horizon agent behavior on the Qwen3.5 architectural foundation. Unlike the hosted Qwen3.8-Max service, these weights are text-only and require thinking mode. The native 256K context can be extended through YaRN. Public weights use the custom Qwen3.8-Max License; hosted Max vision and non-thinking features should not be attributed to this checkpoint.

North Micro Vision Instruct

Cohere

North Micro Vision Instruct is a compact open-weight vision-language model for captioning, visual questions, grounding, and document understanding. It combines a 2B language model with a 400M vision encoder and preserves image aspect ratios through native-resolution processing. The language backbone supports 128K tokens, but multimodal training and validated operating conditions cover only 8K; this page uses the conservative multimodal limit. Apache-licensed weights make it a foundation for domain-specific fine-tuning and local experimentation. It supports interleaved images and text, multilingual prompts, and text output.

LFM2.5 VL 3B

Liquid AI

Liquid’s larger LFM2.5 vision model integrates a 400M SigLIP2 NaFlex encoder with the 2.6B language backbone. It strengthens screen understanding, grounding, multi-image processing, and function calling. The post-trained model answers directly without long reasoning traces, making it suitable for interactive visual assistants. Downloadable weights use the Liquid Open Model License, whose commercial-use conditions should be checked before deployment. Local inference supports private customization, while task reliability and performance require evaluation on the intended hardware.

MAI-Thinking-1

Microsoft

MAI-Thinking-1 is Microsoft AI’s medium-inference-footprint reasoning model, first announced at Build in June and made available in public preview on Microsoft Foundry on August 12, 2026. It uses a sparse mixture-of-experts architecture and was trained in-house without third-party model distillation. The model targets mathematical reasoning, software engineering, and complex enterprise questions. Clean-data and benchmark statements are provider claims rather than guarantees of domain correctness. June announcement-only exposure is not used as its public release date.

Grok 4.6

xAI

Grok 4.6 is a proprietary xAI language-model release. The August frontier update targets reliable coding, agentic workflows, and analytical knowledge work. It supports text and image understanding with textual responses, configurable reasoning, function calling, and structured outputs. Architecture and parameter count remain undisclosed, and task-level improvements do not guarantee factual accuracy. Weights are not publicly downloadable. Published agent and coding results depend on the evaluation harness, tool permissions, and inference settings, so no unsupported standalone benchmark score is assigned.

MAI-Code-1.1-Flash

Microsoft

MAI-Code-1.1-Flash is a substantive update to Microsoft’s small coding model deployed in GitHub Copilot. It improves terminal-agent and .NET tasks and was trained from production feedback with reinforcement-learning environments tied to developer workflows. Microsoft reports improvements in token efficiency, streaming speed, and code retention relative to the June model. These relative production measurements are not interchangeable with standalone benchmark accuracy. The model is accessed through Copilot and does not have publicly downloadable weights.

Nemotron 3.5 Lightning

NVIDIA

Nemotron 3.5 Lightning is an open-weight NVIDIA language-model release. A compact active-parameter reasoning model designed for fast specialized steps in long-running agents. It interleaves Mamba-2, attention, and sparse experts, with speculative decoding options. BF16 weights support customization, while optimized NVFP4 variants target production inference. The license permits commercial and non-commercial use subject to its published terms.

Solar Pro 4

Upstage

Solar Pro 4 is Upstage’s hosted commercial flagship for long-running enterprise agents. It focuses on terminal tasks, multi-turn tool use, document-grounded reasoning, and producing deliverables across multiple files. It supports English, Korean, and Japanese, adjustable reasoning effort, a 512K context, and up to 128K output tokens. The model is offered through Upstage Console and partner APIs rather than open weights. Its August 11 launch blog precedes the August 20 press release. Benchmarks depend on agent environments and do not guarantee task completion in a particular production workflow.

Qwen-Audio-3.0-Realtime

Alibaba

Qwen-Audio-3.0-Realtime is a proprietary language-capable speech conversation model for fluid real-time dialogue. It combines speech reasoning with duplex interaction, allowing conversational responses rather than only transcription or speech synthesis. The 3.0 release offers separately named Plus and Flash serving tiers, grouped here as the audio conversation family. Hosted streaming APIs provide access; no downloadable weights or checkpoint parameter count is verified.

Muse Glimmer

Meta

Muse Glimmer is Meta’s 30B open-weight multimodal model designed for local, always-on agent workflows. It transfers capabilities from Muse Spark through distillation, followed by longer-context mid-training, supervised tuning, reinforcement learning, and on-policy distillation. The model handles text and images, tool calling, coding, and multi-step plans under configurable reasoning effort. Apache-2.0 weights permit private deployment and adaptation. Consumer hardware viability depends on memory and quantization; local operation does not guarantee reliable autonomous behavior.

Ling-3.0-Tiny

InclusionAI (Ant Group)

Ling-3.0-Tiny is a native hybrid reasoning model using Kimi Delta Attention and gated Multi-head Latent Attention in a 3:1 layer pattern. Sparse experts support coding, knowledge work, mathematical reasoning, and tool-based agents, with thinking configurable per request. The Tiny tier targets local and edge use with 7.9B total parameters; the card documents Apple Silicon and DGX Spark deployment. Public API availability precedes the later August weight publication, as established by the official changelog. MIT weights, routine base checkpoints, and quantized distributions are grouped under this tier.

GPT-5.6 Cyber

OpenAI

GPT-5.6 Cyber is the purpose-trained security model identity available to approved defenders through Daybreak Red. It supports advanced authorized vulnerability research, exploit validation, and security testing through Responses. Access requires separate approval and provisioning; it is not an unrestricted public model. The API name is a maintained alias for the program’s advanced cybersecurity models.

SeedRealtime

ByteDance

SeedRealtime is ByteDance’s native full-duplex audio-visual language model for continuous conversation. It unifies text, audio, video, conversational timing, and response generation in one architecture. Streaming perception allows it to use the user’s environment to resolve ambiguous speech and maintain context in noisy multi-speaker scenes. The model was publicly described as fully rolled out in ByteDance products. Future tool-driven actions remain roadmap work; the launch does not establish downloadable weights or a general developer API.

Muse Spark 1.2

Meta

Muse Spark 1.2 is a proprietary Meta multimodal reasoning model for coding and agentic knowledge work. The 1.2 update focuses on coding and tool-driven visual work, launching alongside the Muse Code terminal agent with stronger long-task and multimodal capabilities. It accepts text, images, video, and audio and produces text; the API specifies a 1,048,576-token context for these versions. Standard and, where offered, Contributor access use the same model capabilities but different data-use terms. Contributor access is a serving tier rather than a separate model. Future open-weight plans are not a weight release.

LFM2.5 2.6B

Liquid AI

This dense agent model scales pretraining to about 34T tokens and adds a dedicated 128K context-extension phase. Post-training combines teacher specialization, on-policy distillation, and reinforcement learning for planning and chained tool calls. The model is designed for smartphones, CPUs, and GPU serving, with downloadable base and post-trained checkpoints. Downloadable weights use the Liquid Open Model License, whose commercial-use conditions should be checked before deployment. Local inference supports private customization, while task reliability and performance require evaluation on the intended hardware.

Nemotron Labs VoiceChat

NVIDIA

Nemotron Labs VoiceChat is an open-weight full-duplex voice assistant that combines a streaming Conformer speech encoder, Nemotron Nano v2 language model, and speech decoder. It can listen and speak during an ongoing conversation and emit structured function calls through a separate output channel. The released 11B system is intended for spoken-assistant research and customization, including interaction timing and tool use. The weight card dates public distribution to August 3, 2026; earlier service previews do not establish the availability of this exact downloadable checkpoint.

Qwen3.8-Max

Alibaba

Qwen3.8-Max is the hosted 2.4T-scale flagship targeting long-horizon coding and professional knowledge work. Native visual understanding supports planning, execution, and verification across documents, interfaces, and video. The production service exposes a million-token context, configurable reasoning effort, built-in tools, and non-thinking behavior. Those hosted features differ from the separately downloadable text-only Qwen3.8-2.4T-A95B checkpoint. September serving updates such as Max-0902 are grouped as dated variants.

Ling-3.0-Flash-Fin

InclusionAI (Ant Group)

Ling-3.0-Flash-Fin is Ant Group’s finance-enhanced version of Ling3.0-Flash, developed through continued training with financial institutions and domain experts. It combines source-grounded search, multi-document reasoning, valuation modeling, spreadsheet workflows, and report preparation in long-horizon agents. The 124B model activates 5.1B parameters and supports 256K context. The August announcement predates the September weight publication; the exact first public serving day is not asserted here. MIT weights are now downloadable. Model outputs require review of financial assumptions, calculations, sources, and conclusions.

Intern S2 Mobius

Shanghai AI Laboratory

Intern S2 Mobius is a 35B research model that separates knowledge storage from reasoning computation. Globally shared memory lets recurrent reasoners refine continuous hidden states and access knowledge beyond a fixed layer hierarchy. It is continually pretrained from Qwen3.5-35B and post-trained with supervised learning and reinforcement learning. The public checkpoint and deployment code allow study of concise latent reasoning and multi-token prediction. Primary catalogs verify the release month; the exact first-public day remains unverified and is not inferred from private repository or API creation metadata.

K-EXAONE 2.0

LG AI Research

K-EXAONE 2.0 is LG AI Research’s 750B-parameter sovereign language model with 37B active parameters. Its hybrid-attention mixture-of-experts architecture scales the preceding K-EXAONE release while expanding support to ten languages, including Korean, English, Japanese, and several European languages. Apache 2.0 replaces the earlier generation’s custom licensing, allowing commercial weight deployment. The model is designed for reasoning, coding, and agentic applications with 256K context. Official weights, a technical report, and demo access support reproducible evaluation rather than relying only on the provider’s aggregate benchmark claims.

Gemini Robotics ER 2

Google

Gemini Robotics ER 2 is a public-preview vision-language model for embodied reasoning, based on Gemini 3.5 Flash. It improves spatial reasoning, multi-step tool use, robot-video moment finding and progress classification, and multi-robot coordination. A streaming endpoint adapts the same release to continuous audio and video using the Live API. The model produces text and high-level tool instructions rather than directly serving as a motor controller.

Inkling Small

Thinking Machines Lab

Inkling Small is Thinking Machines Lab’s open-weight multimodal mixture-of-experts model, trained from scratch for native reasoning over text, images, and audio. Variable thinking effort allows applications to choose a latency and reasoning tradeoff. The checkpoint supports up to 1M tokens, while hosted product limits can be lower. Apache-licensed weights and Tinker access enable customization and deployment for coding, conversational assistants, and tool-using agents. The July 15 announcement previewed the Small tier; its full weight release and model card are dated July 30.

Grok Voice Think Fast 2.0

xAI

Grok Voice Think Fast 2.0 is xAI’s reasoning-enabled spoken-dialogue model. It combines low-latency speech interaction with reasoning and tool use for voice assistants. The model can answer spoken questions and take multi-step actions through application tools, so it is distinct from speech transcription or standalone text-to-speech. The second generation improves spoken reasoning, instruction following, and agent tasks over the first version. Its parameter count and detailed architecture are not disclosed. Real-time latency and interaction quality also depend on transport, audio capture, and the surrounding application.

Qwen3.7-Flash

Alibaba

Qwen3.7-Flash is the lower-cost hosted multimodal tier in the Qwen3.7 line. Official documentation confirms text/image/video input, text output, million-token context, function calling, built-in tools, and structured output. It supports high-frequency reasoning, document processing, and agent workloads through a named API model. The provider does not establish a downloadable checkpoint or parameter count for this exact service. The July 25 launch catalog establishes availability; its July 15 serving snapshot is grouped as a variant.

Claude Opus 5

Anthropic

Claude Opus 5 is a major Opus-generation release for efficient coding, knowledge work, computer use, and scientific research. It emphasizes proactive problem solving, verifying results, and repairing root causes while approaching Fable 5 capability at lower cost. Its million-token context and adaptive thinking support sustained professional workflows, with text and image inputs and tools through the Claude API and partner platforms.

Ling-3.0-Flash

InclusionAI (Ant Group)

Ling-3.0-Flash is a native hybrid reasoning model using Kimi Delta Attention and gated Multi-head Latent Attention in a 5:1 layer pattern. Sparse experts support coding, knowledge work, mathematical reasoning, and tool-based agents, with thinking configurable per request. The Flash tier targets production agent efficiency with 124B total parameters. Public API availability precedes the later August weight publication, as established by the official changelog. MIT weights, routine base checkpoints, and quantized distributions are grouped under this tier.

Solar Open 2

Upstage

Solar Open 2 is Upstage’s 250B-parameter open-weight model for document-heavy agents, office productivity, and coding. It activates 15B parameters per token and interleaves linear attention with standard attention to improve long-context inference efficiency. A 1M-token context supports large document collections and extended tool histories. Its release includes weights and a technical report describing training and evaluation. The custom Solar Open license governs redistribution and commercial use. Sparse activation lowers computation relative to a dense model of the same total size but does not eliminate the storage and memory requirements of the full weights.

Gemini 3.5 Flash-Lite

Google

Gemini 3.5 Flash-Lite is a generally available efficiency tier for low-latency automation and subagents. It combines a million-token input context with multimodal understanding, thinking, structured outputs, and native tools. Its positioning emphasizes cost-effective high-volume work rather than the deepest reasoning. Computer use is supported as a preview capability and needs a compatible execution environment.

Gemini 3.6 Flash

Google

Gemini 3.6 Flash is a generally available update emphasizing token efficiency, more concise responses, and stronger coding and agentic planning. It maintains multimodal input and long-context text output while improving the cost profile over 3.5 Flash. Thinking levels and native tools support sustained workflows. Sampling controls such as temperature, top-p, and top-k are deprecated for this generation.

Intern S2 Preview 397B

Shanghai AI Laboratory

Intern S2 Preview scales multimodal pretraining, scientific reinforcement learning, and long-horizon agent environments. It learns from raw pages of scientific literature and combines general reasoning with tasks across more than twenty scientific domains. The 397B checkpoint supports language output, image understanding, and tool calls. Its text evaluations use up to 256K inference, while multimodal evaluations use shorter 64K limits. The official laboratory announcement identifies the public release date as 2026-07-17.

Kimi K3

Moonshot AI

Kimi K3 is a 2.8-trillion-parameter native multimodal agent model with a million-token context. Its architecture combines Kimi Delta Attention, Gated Multi-head Latent Attention, Attention Residuals, and Stable LatentMoE, activating 104B parameters per token. Quantization-aware training uses MXFP4 weights and MXFP8 activations. It supports long-running coding, knowledge-work, and visually grounded agent tasks through Kimi, Kimi Work, Kimi Code, and the official API. Full weights are now available under the custom Kimi K3 License; the initial July 16 release provided hosted access before the weight release.

Inkling

Thinking Machines Lab

Inkling is Thinking Machines Lab’s open-weight multimodal mixture-of-experts model, trained from scratch for native reasoning over text, images, and audio. Variable thinking effort allows applications to choose a latency and reasoning tradeoff. The checkpoint supports up to 1M tokens, while hosted product limits can be lower. Apache-licensed weights and Tinker access enable customization and deployment for coding, conversational assistants, and tool-using agents. Its pretraining includes 45T tokens spanning text, images, audio, and video; video training does not establish a supported direct video-input endpoint.

Bonsai 27B

PrismML

Bonsai 27B is PrismML’s larger multimodal tier, derived from Qwen3.6-27B and optimized for local devices. It retains a hybrid linear/full-attention language backbone and image understanding, while binary language weights sharply reduce storage and memory needs. The model supports a 262K context, with an optional vision-tower package for image inputs; language-only deployment can omit that package. Apache-licensed weights are provided in GGUF and MLX formats, alongside ternary variants. July 14 is the official public release date. Hardware demonstrations and packed-weight footprints do not remove the additional memory cost of context, activation buffers, and runtime components.

Agents-A1 4B

Shanghai AI Laboratory

Agents-A1 4B is Shanghai AI Laboratory’s model for long-horizon search, engineering, and scientific agent workflows. Its training combines long verified trajectories, domain experts, and on-policy distillation to consolidate specialized skills into one language model. It generates reasoning, text, and tool requests; external tools and a harness provide execution. Official cards explicitly date the 2026-07-14 release and document a 262K context. Apache-licensed weights and evaluation code support local deployment, while benchmark performance depends on the evaluation environment and external search infrastructure.

Muse Spark 1.1

Meta

Muse Spark 1.1 is a proprietary Meta multimodal reasoning model for coding and agentic knowledge work. The 1.1 update materially improves repository coding, multi-turn agents, and multimodal workflows, and introduces the public Meta Model API preview. It accepts text, images, video, and audio and produces text; the API specifies a 1,048,576-token context for these versions. Standard and, where offered, Contributor access use the same model capabilities but different data-use terms. Contributor access is a serving tier rather than a separate model. Future open-weight plans are not a weight release.

GPT-5.6 Luna

OpenAI

GPT-5.6 Luna is the cost-oriented GPT-5.6 tier, occupying the role of earlier Nano models. It combines text and image input with a long API context and reasoning effort settings from none to max. It is intended for focused, repeated tasks and as an economical worker in larger workflows; more effort can increase both latency and output consumption.

GPT-5.6 Sol

OpenAI

GPT-5.6 Sol is the flagship tier of the GPT-5.6 family; the unsuffixed gpt-5.6 alias resolves to this model. It supports professional coding and reasoning with a broad effort range including max. The release adds persisted reasoning, programmatic tool calling, explicit caching, and multi-agent API orchestration, allowing extended tool-assisted tasks with text and image input.

GPT-5.6 Terra

OpenAI

GPT-5.6 Terra is the middle GPT-5.6 tier for balancing intelligence and cost, occupying the role previously served by Mini models. It accepts text and images and offers reasoning effort from none through max. Its long-context API and GPT-5.6 workflow features make it suitable for coding workers and repeated agent tasks where flagship cost is difficult to justify.

UltraX

OpenBMB

UltraX is OpenBMB’s specialized 0.6B language-model release for refining pre-training text through structured editing function calls. It generates insertions, deletions, and replacements rather than rewriting an entire document, and a deterministic executor applies validated edits. Full-parameter supervised training uses program supervision built from language-model refinements and alignment procedures. The July 9 paper precedes the July 13 public code-and-model announcement. Apache-licensed refinement checkpoints and data are available, with no-instruction and preservation-weighted variants grouped here. Its named language-editing specialization qualifies separately from generic data tools or probability-only classifiers.

SWE-1.7

Cognition

SWE-1.7 is Cognition’s Kimi K2.7 Code-derived model for long-running software-engineering tasks, released through Devin Web, Desktop, and CLI. Extensive reinforcement learning targets thorough codebase exploration, faithful task execution, and concise reasoning while preserving the ability to think longer on hard problems. Learned self-compaction summarizes working state and resumes from those summaries, allowing agent sessions to continue beyond one model context window without implying an unlimited single-call context. Training uses hardened executable environments and alternating cost penalties. The publisher explicitly distinguishes its behavior from the base checkpoint. Cognition does not publish weights or verified standalone model limits for this release.

Grok 4.5

xAI

Grok 4.5 is a proprietary xAI language-model release. A frontier model for software engineering, multi-step agents, and knowledge work. It supports image and text inputs, configurable reasoning effort, function calling, and structured outputs. The API changelog dates availability to July 8, before the July 16 public announcement. There is no stated text-output token cap in the current endpoint documentation. Weights are not publicly downloadable. Published agent and coding results depend on the evaluation harness, tool permissions, and inference settings, so no unsupported standalone benchmark score is assigned.

Granite SWASH

IBM

Granite SWASH is IBM Research’s open architectural preview for efficient language modeling. The family explores sliding-window attention at every layer, learnable attention sinks, and hybrid transformer designs, with dense 2B and sparse 3B checkpoints. The sparse model activates approximately 600M parameters. These are general-purpose English base models for research and customization, rather than instruction-ready assistants. Apache-licensed weights allow study of the architecture and downstream training; chat and tool behavior should not be inferred from the base release.

Nemotron Labs 3 Puzzle

NVIDIA

Nemotron Labs 3 Puzzle is an open-weight NVIDIA language-model release. Puzzle is a deployment-oriented structural compression of Nemotron 3 Super. It reduces the teacher’s parameter and expert budget while retaining its hybrid architecture, reasoning behavior, and multi-token prediction. NVIDIA releases BF16, FP8, and NVFP4 checkpoints for different serving systems. The license permits commercial and non-commercial use subject to its published terms.

Nemotron Labs Audex

NVIDIA

Audex is NVIDIA’s unified audio-text language-model family, designed to add speech and sound understanding and generation while retaining text reasoning. Audio encoders project inputs into the language embedding space; quantized audio and text tokens share a decoder. The main 30B sparse checkpoint builds on Nemotron-Cascade 2, and a smaller dense 2B tier is also released. It supports audio question answering, spoken dialogue, speech translation, and text reasoning. The research paper publicly releases the checkpoints on July 6; the card’s June 8 date conflicts with that publication record.

GPT-Realtime-2.1 Mini

OpenAI

GPT-Realtime-2.1 Mini is OpenAI’s realtime conversational model for low-latency speech and text interaction. Its bidirectional Realtime API supports voice-agent sessions, image input, and external tools. The smaller tier reduces cost and response latency for repeated conversational work. The 2.1 update improves alphanumeric recognition, noise and silence handling, and interruption behavior. Configurable reasoning supports more complex voice workflows.

GPT-Realtime-2.1

OpenAI

GPT-Realtime-2.1 is OpenAI’s realtime conversational model for low-latency speech and text interaction. Its bidirectional Realtime API supports voice-agent sessions, image input, and external tools. The 2.1 update improves alphanumeric recognition, noise and silence handling, and interruption behavior. Configurable reasoning supports more complex voice workflows.

Hy3

Tencent

Hy3 is Tencent’s stable follow-up to Hy3 Preview, trained with higher-quality post-training data and feedback from deployed products. It retains a 295B-parameter sparse architecture with 21B active parameters and a 256K context. The model supports both fast responses and slower reasoning, with improvements to tool use, coding, and productivity workflows. Apache-licensed weights replace the preview’s custom community license. Official API access and integration into Tencent products provide hosted alternatives to private deployment.

LLaDA2.2

InclusionAI (Ant Group)

LLaDA2.2 is Ant Group’s MoE diffusion-language model family, generating text through block-parallel refinement rather than exclusively left-to-right decoding. Levenshtein editing introduces DELETE and INSERT tokens for structural correction, while block routing and agentic reinforcement learning support 128K tool use and multi-turn interaction. The 100B Flash checkpoint was followed by the 16B Mini release in September. Public Apache-2.0 weights and custom inference code support research and self-hosting. Performance and latency depend on diffusion settings and the serving implementation.

Sarvam 105B Conversations

Sarvam AI

Sarvam released updated 105B conversational and work-agent variants at its July 2026 Epoch event. The conversational specialization targets low-latency agents, noisy speech transcripts, code-mixed language, policy following, and multi-turn tool calls. It generates text as part of a voice-agent stack rather than natively replacing speech recognition or synthesis. Both are documented as available through the Sarvam API. Public downloadable weights and full endpoint specifications for these updated checkpoints are not established. The release month is retained because the two-day event does not establish an exact first-availability day.

Sarvam 105B Work

Sarvam AI

Sarvam released updated 105B conversational and work-agent variants at its July 2026 Epoch event. The work-agent specialization is a reasoning variant tuned for sandboxed software engineering, data analysis, knowledge retrieval, office tasks, and long-running autonomy. It is distinct from the original open checkpoint and from Indus’s routing to third-party models. Both are documented as available through the Sarvam API. Public downloadable weights and full endpoint specifications for these updated checkpoints are not established. The release month is retained because the two-day event does not establish an exact first-availability day.

Claude Sonnet 5

Anthropic

Claude Sonnet 5 is the next-generation Sonnet tier, intended to narrow the gap to Opus 4.8 on coding and agent work at a lower cost. It plans multi-step jobs, uses browsers and terminals, and checks results more consistently than Sonnet 4.6. Text and image inputs, adaptive thinking, and a million-token context support both everyday assistance and longer development workflows.

Leanstral 1.5

Mistral AI

Leanstral 1.5 is Mistral’s open-weight specialist model for Lean 4 formal mathematics and program verification. The 1.5 update improves proof engineering through mid-training, supervised tuning, and reinforcement learning in theorem-proving and code-agent environments. Its first API release is dated June 30 in the changelog, before the July 2 announcement. It edits proof files, runs tools, and revises proofs after compiler feedback within an agent scaffold. Apache-2.0 weights permit local deployment and adaptation. Published success rates involve explicit inference budgets and multiple attempts; a long agent run is not the same as a single context window.

Audio Visual Flamingo

NVIDIA

Audio Visual Flamingo is an open-weight NVIDIA audio-language research model. AVF combines audio and video evidence with textual instructions for synchronized event understanding and natural-language answers. It uses a Qwen2.5-7B backbone, audio front-end, and vision components derived from NVIDIA’s multimodal research. The final released checkpoint is dated June 29, 2026; earlier repository creation is not used as evidence of public availability. The release supports research use under a non-commercial license. Descriptions of perceptual content are generated language and can misidentify events, speakers, or musical attributes.

Agents-A1

Shanghai AI Laboratory

Agents-A1 is Shanghai AI Laboratory’s model for long-horizon search, engineering, and scientific agent workflows. Its training combines long verified trajectories, domain experts, and on-policy distillation to consolidate specialized skills into one language model. It generates reasoning, text, and tool requests; external tools and a harness provide execution. Official cards explicitly date the 2026-06-26 release and document a 262K context. Apache-licensed weights and evaluation code support local deployment, while benchmark performance depends on the evaluation environment and external search infrastructure.

LFM2.5 230M

Liquid AI

Liquid’s smallest LFM2.5 text tier is designed for inexpensive local inference and downstream fine-tuning. Its training includes 19T tokens and a 32K context extension, followed by distillation, preference optimization, and reinforcement learning. It can select tools from natural-language instructions, but demonstrations of robot control rely on external low-level skills. Downloadable weights use the Liquid Open Model License, whose commercial-use conditions should be checked before deployment. Local inference supports private customization, while task reliability and performance require evaluation on the intended hardware.

Nemotron Labs TwoTower

NVIDIA

Nemotron Labs TwoTower is an open research language model that separates autoregressive context processing from block-wise diffusion generation. A frozen Nemotron 3 Nano context tower encodes preceding clean tokens, while a trainable denoiser tower refines blocks through cross-attention. The released system contains both sparse towers. It investigates parallel token generation and the quality-throughput tradeoff rather than a new chat product. Reported speedups depend on block size, confidence thresholds, hardware, and the autoregressive baseline used in the paper.

Qwen-AgentWorld

Alibaba

Qwen-AgentWorld is an open language world model for simulating agent environments through reasoning and next-observation prediction. The 35B-A3B checkpoint covers MCP, search, terminal, software engineering, Android, web, and operating-system interactions using language representations. Continued pretraining, supervised training, and reinforcement learning on real interaction trajectories target faithful environment simulation. The technical report appeared June 23, followed by the June 24 weight announcement. This is a distinct language-model specialization rather than a tool-execution application.

Doubao Seed 2.1 Pro

ByteDance

Doubao Seed 2.1 Pro is ByteDance’s flagship productivity model in the Seed 2.1 family. It combines language reasoning with image and video understanding for document-heavy tasks, visual analysis, and multi-step agents. Public inference is offered through Volcengine and ByteDance applications, with no downloadable weights. The tier is a separate product identity; API snapshots and later service migrations are variants. Parameter counts and unsupported endpoint limits are left unspecified.

Doubao Seed 2.1 Turbo

ByteDance

Doubao Seed 2.1 Turbo is ByteDance’s efficient productivity tier in the Seed 2.1 family. It combines language reasoning with image and video understanding for document-heavy tasks, visual analysis, and multi-step agents. Public inference is offered through Volcengine and ByteDance applications, with no downloadable weights. The tier is a separate product identity; API snapshots and later service migrations are variants. Parameter counts and unsupported endpoint limits are left unspecified.

GLM-5.2

Z.ai

GLM-5.2 expands the GLM engineering line to a million-token context and improves reliability on sustained coding tasks. IndexShare reuses sparse-attention indexing across groups of layers to lower long-context computation, while multi-token prediction improvements accelerate decoding. Training emphasizes maintaining goals and using noisy repository and tool context over extended sessions. Public MIT weights and official API access support deployment. The million-token limit is an input-context capability, not a guarantee of reliable retrieval or correct execution at every length. GLM-5.2 Fast is a faster-serving route with the same weights, grouped here rather than cataloged as a separate trained model.

Kimi K2.7 Code

Moonshot AI

Kimi K2.7 Code is a separately positioned coding specialization of the native multimodal K2 family. It improves long-horizon software engineering and reasoning efficiency while retaining image and video input for visual coding tasks. The model supports a 256K context and always runs in thinking mode; Kimi Code requests with thinking disabled fall back to K2.6. Public modified-MIT weights support self-hosting, and the official kimi-k2.7-code API exposes the model to coding agents. Its identity is distinct from a general-purpose K2.7 release, which is not implied by this page.

DiffusionGemma

Google

DiffusionGemma is an experimental open-weight text diffusion model built on the Gemma 4 26B A4B backbone. An autoregressive encoder caches context while a bidirectional decoder repeatedly denoises 256-token canvases, producing blocks of text in parallel. It accepts text and visual input and supports thinking and tools. It targets fast small-batch generation on a capable accelerator, trading some task accuracy for throughput relative to autoregressive Gemma 4.

Claude Fable 5

Anthropic

Claude Fable 5 is the publicly accessible Mythos-class Claude model for extended software engineering, visual reasoning, knowledge work, and research. It uses adaptive thinking and can maintain progress through tool-driven tasks with file-based memory. Claude Mythos 5 is the same underlying model with different safeguards and restricted trusted access, represented here as a variant. Fable access was suspended June 12 and restored July 1; the first release remains June 9.

North Mini Code

Cohere

North Mini Code is Cohere’s first North coding model, trained for repository-level software changes and terminal-based agent tasks. Its mixture-of-experts architecture has 30B total parameters and 3B active per token, allowing relatively efficient inference while retaining the memory needs of the full model. Training across multiple agent harnesses targets transferable software engineering behavior rather than a single fixed scaffold. Apache-licensed weights, Chat V2 access, and private Model Vault deployments are available. The documented context limit is 256K with up to 64K output tokens.

AFM 3 Cloud Pro

Apple

AFM 3 Cloud Pro is a third-generation Apple Foundation Model published at WWDC26 as part of Apple Intelligence. The highest-capability server tier is built with Google and runs on NVIDIA GPUs through an extension of Private Cloud Compute for demanding reasoning and agentic tool use. Access is integrated into Apple operating-system features rather than provided as freely downloadable weights. Availability depends on the device, software version, and supported feature. The June 8 date records the published generation; it does not imply that every consumer feature became generally available that day.

AFM 3 Cloud

Apple

AFM 3 Cloud is a third-generation Apple Foundation Model published at WWDC26 as part of Apple Intelligence. The server workhorse advances Apple’s PT-MoE architecture for multimodal reasoning while running on Apple silicon in Private Cloud Compute. Access is integrated into Apple operating-system features rather than provided as freely downloadable weights. Availability depends on the device, software version, and supported feature. The June 8 date records the published generation; it does not imply that every consumer feature became generally available that day.

AFM 3 Core Advanced

Apple

AFM 3 Core Advanced is a third-generation Apple Foundation Model published at WWDC26 as part of Apple Intelligence. Its prompt-level expert selection loads a subset of weights from flash into DRAM, enabling a larger local multimodal model within constrained memory. Access is integrated into Apple operating-system features rather than provided as freely downloadable weights. Availability depends on the device, software version, and supported feature. The June 8 date records the published generation; it does not imply that every consumer feature became generally available that day.

AFM 3 Core

Apple

AFM 3 Core is a third-generation Apple Foundation Model published at WWDC26 as part of Apple Intelligence. The compact dense model is optimized for Apple silicon and local everyday text tasks. Access is integrated into Apple operating-system features rather than provided as freely downloadable weights. Availability depends on the device, software version, and supported feature. The June 8 date records the published generation; it does not imply that every consumer feature became generally available that day.

Nemotron 3 Ultra

NVIDIA

Nemotron 3 Ultra is an open-weight NVIDIA language-model release. The largest released Nemotron 3 tier targets difficult reasoning, software engineering, and long-running tool workflows. Latent mixture-of-experts routing and multi-token prediction support efficient frontier-scale training and serving. Downloadable base and post-trained weights are available in several precisions. The license permits commercial and non-commercial use subject to its published terms.

Gemma 4 12B

Google

Gemma 4 12B is a separately released multimodal update using a unified, encoder-free architecture. Image patches and audio waveforms enter through linear projections into the language model instead of dedicated modality encoders. Its approximately 12B dense decoder supports text, vision, audio, thinking, and native tool use with a 256K context. Quantized deployment targets consumer hardware, including laptops with 16GB of memory; actual capacity depends on precision and context.

Holo3.1

H Company

Holo3.1 is H Company’s updated computer-use model family, adding mobile robustness, native function-calling protocols, and practical local deployment. It spans dense 0.8B, 4B, and 9B Qwen3.5-derived tiers plus a sparse 35B-A3B tier whose current card identifies Qwen3.6 ancestry. The release includes Apache-licensed weights and optimized FP8, NVFP4, and Q4 GGUF options for the larger tier. These are variants of the same trained generation rather than extra catalog models. The current H Models API exposes the 35B tier with a 65,536-token hosted context; that limit is described without treating it as a verified uniform specification for every size.

MAI-Code-1-Flash

Microsoft

MAI-Code-1-Flash is Microsoft AI’s small active-parameter coding model, deployed in GitHub Copilot for everyday developer work. It was trained together with the production Copilot harness to handle repository questions, refactoring, multi-turn code changes, and adaptive reasoning. Microsoft reports five billion active parameters but has not disclosed the full architecture or total size. The model remains proprietary; access in a developer product is distinct from downloadable weights or a generally available independent API.

MiniMax M3

MiniMax

MiniMax M3 is a native multimodal MoE model trained jointly on text, images, and video. It adds MiniMax Sparse Attention for efficient million-token context and targets long-horizon coding and knowledge-work agents. The model supports enabled, adaptive, and disabled reasoning modes. MiniMax launched hosted API and Agent access on June 1 and later released downloadable weights. The full checkpoint has about 428B parameters with approximately 23B activated. Its MiniMax Community License permits non-commercial use and imposes additional conditions for commercial deployment.

Nex-N2

Nex AGI

Nex-N2 is Nex AGI’s multimodal agentic model family, with a 35B mini checkpoint based on Qwen3.5-35B-A3B-Base and a 397B Pro checkpoint based on Qwen3.5-397B-A17B. Post-training develops agentic thinking that balances direct answers with sustained reasoning and tool execution. The family targets coding, computer interaction, research, and knowledge workflows. Apache-licensed weights and hosted partner access are documented. June availability is established by the official model card’s June 9 hosted-access promotion; first public day is unresolved, so the release month is retained. Routine precision variants remain grouped with their checkpoints.

Qwen3.7-Plus

Alibaba

Qwen3.7-Plus is a hosted multimodal tier for coding, professional work, and long-horizon agents. It accepts text, images, and video, returns text, and supports a million-token context. Official API documentation lists function calling, built-in tools, structured output, and thinking controls. It is separately positioned from the earlier Qwen3.7-Max and the lower-cost Flash tier. No downloadable weights or checkpoint-level parameter count are established for this exact hosted identity.

Cosmos 3 Super Reasoner

NVIDIA

Cosmos 3 Super Reasoner is the language-generating 32B reasoning tower of NVIDIA’s Cosmos 3 Super world model, adapted from the Qwen3-VL 32B architecture. It accepts text, images, and video and produces textual reasoning about physical scenes, spatial relationships, trajectories, and visual events with a documented 256K context. NVIDIA’s downloadable Cosmos3-Super checkpoint packages this autoregressive tower with a separate 32B diffusion generator, making the full omnimodal checkpoint 64B rather than a 64B language reasoner. NIM exposes the reasoning tower as nvidia/cosmos3-super-reasoner; Nano’s 8B reasoner is a smaller family variant. OpenMDW 1.1 permits commercial and non-commercial model use.

Step 3.7 Flash

StepFun

Step 3.7 Flash is an open-weight vision-language model for high-throughput agent execution. Its 198B total parameters combine a 196B sparse language backbone with a 1.8B vision encoder, activating about 11B parameters per token. It accepts images and text, supports a 256K context, and offers low, medium, and high reasoning levels. Training targets visual search, browser and terminal tool use, coding, and professional document workflows. API, web chat, and downloadable weights are available. Advisor-mode results depend on an additional larger model and should not be interpreted as standalone checkpoint performance.

Claude Opus 4.8

Anthropic

Claude Opus 4.8 is a further independent Opus update for sustained coding, agent tasks, and knowledge work. Anthropic emphasizes more dependable judgment, honesty about incomplete work, and stronger follow-through. It provides a million-token context by default and adaptive thinking, with the same tool surface as Opus 4.7. Its fast serving mode is an access option rather than a separate checkpoint.

LFM2.5 8B A1B

Liquid AI

The sparse LFM2.5 tier combines expert routing, short convolutions, and attention for local agents. It extends context to 128K and expands the tokenizer for non-Latin scripts. Unlike its preceding LFM2 checkpoint, the post-trained version always generates reasoning traces. The model card’s measured parameter counts are more precise than the rounded A1B product name. Downloadable weights use the Liquid Open Model License, whose commercial-use conditions should be checked before deployment. Local inference supports private customization, while task reliability and performance require evaluation on the intended hardware.

Mistral Medium 3.5

Mistral AI

Mistral Medium 3.5 is a dense open-weight multimodal model that combines instruction following, coding, and adjustable reasoning in a single checkpoint. It launched in public preview as the default model in Le Chat and Mistral Vibe, supporting long-running software and productivity tasks, parallel tool calls, and structured outputs. A newly trained vision encoder handles variable image dimensions and aspect ratios. The modified MIT license must be checked for commercial restrictions; the published four-GPU deployment is a configuration example rather than a universal memory requirement.

Hy MT2

Tencent

Hy MT2 is Tencent’s second-generation multilingual translation language model family. It includes 1.8B and 7B dense models and a 30B sparse tier with approximately 3B active parameters. The models support 33 languages and five regional languages or dialects, with terminology constraints, contextual translation, and improved instruction following. Apache-licensed weights and a translation application enable local and hosted evaluation. Quantized mobile builds are variants of the released language checkpoints; they are not separate language-model identities.

Command A+

Cohere

Command A+ unifies Cohere’s enterprise reasoning, retrieval, translation, and image understanding in a sparse mixture-of-experts model. It has 218B total parameters with 25B active per token, supports 48 languages, and combines text and image inputs with text generation. Unlike earlier Command A checkpoints licensed for non-commercial weight use, its downloadable weights use Apache 2.0. Hosted Chat access and private deployment support enterprise agents that must reason over documents and call external tools. Its 128K context and 64K output limits are the documented product limits.

Qwen3.5-LiveTranslate-Flash

Alibaba

Qwen3.5-LiveTranslate-Flash is a language-capable real-time interpretation model that uses speech and visual context to translate conversations. It recognizes 60 source languages and supports spoken output in 29 languages, with text and audio responses exposed through the Realtime API. Visual grounding helps resolve terminology and references during meetings, classes, and livestreams. It builds on Qwen3.5-Omni and adds voice cloning and terminology handling. Weights are not publicly downloadable.

Qwen3.7-Max

Alibaba

Qwen3.7-Max is a proprietary agent foundation model for coding, office automation, and sustained autonomous execution. Training and evaluation emphasize long-horizon tasks, cross-harness generalization, MCP integrations, and multi-step software engineering. Qwen provides the model through Alibaba Cloud Model Studio APIs. It supports million-token contexts and controllable reasoning effort; weights are not released. The later multimodal Plus and Flash tiers are separate hosted identities rather than aliases for this text-focused release. The June 10 serving update adds native visual understanding, represented as a dated variant.

Gemini 3.5 Flash

Google

Gemini 3.5 Flash is a generally available model update for sustained coding and agentic work. It combines multimodal understanding with a million-token input context, thinking, and tools for code execution and grounded retrieval. Computer use was subsequently exposed as a public preview tool for browser, mobile, and desktop workflows. The Antigravity managed agent uses this model but is a separate application.

Nemotron Labs Diffusion

NVIDIA

Nemotron Labs Diffusion is a tri-mode research family that unifies autoregressive generation, diffusion decoding, and self-speculative decoding in one trained language model. Different attention patterns and decoding procedures let users explore generation speed and output quality with a shared checkpoint. Released 3B, 8B, and 14B language tiers include base and instruction models; a related 8B vision-language checkpoint adds image input. The research targets inference efficiency and hybrid decoding, and measured throughput is specific to the chosen sampler and hardware.

MiniCPM5

OpenBMB

MiniCPM5 is OpenBMB’s fifth-generation dense language-model family for local and resource-constrained use. The 1.08B model was announced May 19, followed by a 2.52B model on September 7. Both support a documented 128K context and target reasoning, coding, tool use, writing, and instruction following. The training recipe combines high-quality data, instruction tuning, and complementary reinforcement-learning signals. Apache-licensed base, intermediate, supervised, and final checkpoints are grouped as variants, alongside routine deployment formats. Small parameter counts improve local practicality but do not eliminate errors or the need for application-specific evaluation.

Grok Build 0.1

xAI

Grok Build 0.1 is a proprietary xAI language-model release. A coding model trained specifically for long agentic software-engineering workflows. It is distinct from the Grok Build client application: the released model has its own API slug and training specialization. The developer changelog records early access on May 19, followed by the broader May 29 announcement. Architecture, parameter count, and a stable context specification are not disclosed in the cited release. Weights are not publicly downloadable. Published agent and coding results depend on the evaluation harness, tool permissions, and inference settings, so no unsupported standalone benchmark score is assigned.

Composer 2.5

Cursor

Composer 2.5 is Cursor’s newer Kimi K2.5-derived coding model, significantly improving sustained agent work, tool selection, instruction following, and effort calibration. Training scales synthetic reinforcement-learning tasks and adds localized textual feedback that teaches desired behavior through on-policy distillation. The model retains a documented 200K context window, with standard and fast serving options representing the same intelligence tier. It is available within Cursor’s desktop, web, CLI, and mobile surfaces; the provider also describes building agents with its SDK. The larger model being trained with SpaceXAI is a future project, not a released Composer 3 identity. Composer 2.5 weights remain proprietary.

Intern-S2 Preview 35B

Shanghai AI Laboratory

Intern-S2 Preview 35B is the efficient scientific multimodal preview in Shanghai AI Laboratory’s next generation. Continued pretraining from Qwen3.5 and scientific task scaling strengthen reasoning and agent use while activating about 3B of 35B parameters. Specialized real-valued prediction modules extend the model to scientific structure tasks alongside natural-language and image understanding. The laboratory explicitly dates the weight release to May 15. This identity remains available separately from the later 397B preview; the legacy intern-s2-preview API alias currently points to this 35B model.

MiniCPM-V 4.6

OpenBMB

MiniCPM-V 4.6 is OpenBMB’s compact 1.3B vision-language model for efficient device deployment. It combines Qwen3.5-0.8B with SigLIP2-400M and mixed 4x/16x visual-token compression, accepting text, images, and video and producing language answers. The release focuses on reducing visual encoding work while preserving multi-image and video understanding. Apache-licensed weights, ordinary and Thinking checkpoints, a demo, and a later May 17 API are available. This page does not equate the backbone configuration maximum with validated long-context visual use. Throughput comparisons are provider measurements tied to particular input and serving conditions.

ERNIE 5.1

Baidu

ERNIE 5.1 is Baidu’s more efficient agent and reasoning model derived from ERNIE 5.0’s elastic sub-network training. It compresses total and active parameter capacity while applying asynchronous reinforcement learning and multi-teacher distillation. The hosted model targets language reasoning, creative writing, search, and tool-driven workflows. The official May 9 launch post confirms chat and AI Studio access. Absolute parameter counts and a supported context ceiling are omitted because the announcement establishes relative reductions rather than exact specifications.

Ring-2.6-1T

InclusionAI (Ant Group)

Ring-2.6-1T is Ant Group’s thinking flagship for demanding reasoning and sustained agent workflows. The trillion-parameter hybrid-attention MoE model supports adjustable high and xhigh reasoning effort, allowing developers to select between routine multi-turn agents and more expensive mathematical or scientific reasoning. Asynchronous reinforcement learning and IcePop improve training stability and long-horizon execution. The official May 9 launch precedes the May 15 weight/API announcement. MIT weights are public; native 128K context extends to 256K with YaRN, and tool execution requires an external agent environment.

EMO

Allen Institute for AI

EMO is Ai2’s open mixture-of-experts language model trained so reusable domain modules emerge without predefined semantic labels. Its 128 routed experts activate eight per token, giving 14B total and about 1B active parameters. A document-level routing objective encourages coherent expert groups that can be selected for particular tasks. The main checkpoint is trained on one trillion tokens with additional annealing. Ai2 releases weights, matched baselines, and training code for modularity research. Results for reduced expert subsets depend on the selection method and evaluation task.

Nemotron Labs 3 Elastic

NVIDIA

Nemotron Labs 3 Elastic is an open-weight NVIDIA language-model release. Star Elastic packages several nested reasoning models in a single checkpoint derived from Nemotron 3 Nano. Importance-based component ranking and elastic post-training let developers extract smaller subnetworks without independently training each size. The 23B and 12B variants share the larger checkpoint’s parameter space. The license permits commercial and non-commercial use subject to its published terms.

GPT-Realtime-2

OpenAI

GPT-Realtime-2 is OpenAI’s realtime conversational model for low-latency speech and text interaction. Its bidirectional Realtime API supports voice-agent sessions, image input, and external tools. The second generation adds configurable reasoning and stronger instruction following for complex voice workflows.

Granite Switch 4.1 Preview

IBM

Granite Switch 4.1 Preview adds explicit switching between ordinary language behavior and specialized programmatic capabilities. It is designed to combine a general model with deterministic functions through IBM’s Granite Libraries and Mellea tooling. The release includes 3B, 8B, and 30B language checkpoints under Apache 2.0, with 128K context for the representative 30B model. This is a publicly released preview, distinct from the ordinary Granite 4.1 family. Applications must supply the associated runtime to use specialized functions reliably.

Granite 4.1

IBM

Granite 4.1 is IBM’s enterprise language model release focused on instruction following and tool calling without requiring long reasoning traces. The language family offers 3B, 8B, and 30B checkpoints, trained with staged data refinement and released under Apache 2.0. This page represents the language checkpoints; vision, speech, embedding, and guardrail products in the same collection have different scopes. The models support local and managed deployment, allowing organizations to build private assistants and agent workflows around documented model cards rather than relying on an opaque hosted-only service.

Granite Vision 4.1

IBM

Granite Vision 4.1 is IBM’s 4B vision-language model for technical document understanding. It generates language answers from images, charts, tables, and diagrams, combining a compact language backbone with a vision encoder. The release targets enterprise workflows where users need grounded extraction and interpretation, rather than image generation. Apache-licensed weights support private deployment. Its model card documents task-specific evaluation and deployment instructions; results on chart benchmarks should not be treated as guaranteed accuracy on arbitrary business documents.

Ling-2.6-1T

InclusionAI (Ant Group)

Ling-2.6-1T is Ant Group’s general flagship for coding, long-document tasks, and multi-step agents. A hybrid MLA/linear-attention trillion-parameter MoE model targets efficient inference and concise fast thinking. Process-redundancy suppression during post-training lowers unnecessary output tokens, while agentic training improves tool calling, coding, and instruction execution. The April 29 official changelog establishes API and chat availability before the April 30 open-weight announcement. Public MIT weights support self-hosting; the model card demonstrates 256K deployment, while hosted services can expose different limits.

Holotron 3 Nano

H Company

Holotron 3 Nano is H Company’s Nemotron 3 Nano Omni-derived vision-language policy for computer-use agents. The hybrid Mamba/attention mixture-of-experts backbone has a 30B language tier with about 3B active per token and a native 262,144-token configuration. Post-training on proprietary agent data improves multi-step visual workflows and interface grounding, while the vision encoder supports varied screenshot resolutions. The April 28 release provides downloadable weights under NVIDIA model terms and made the model available in HoloTab. This record covers text and image understanding explicitly described for the released agent model, without assuming every audio/video capability of its Omni backbone.

Nemotron 3 Nano Omni

NVIDIA

Nemotron 3 Nano Omni combines a Nemotron 3 Nano language backbone with image and audio encoders for unified multimodal reasoning. It accepts text, images, audio, and video and produces textual answers, summaries, and tool calls. The model targets document understanding, video question answering, and spoken-input agents while avoiding a separate model for each input type. The backbone has approximately 31B parameters with about 3B active; encoder parameters and multimodal tokens also contribute to memory and context usage.

DeepSeek-V4-Flash

DeepSeek

DeepSeek-V4-Flash is the smaller, efficiency-focused tier of the V4 family, first publicly released as a preview on April 24, 2026. Its hybrid Compressed Sparse Attention and Heavily Compressed Attention architecture supports million-token contexts alongside manifold-constrained hyper-connections. It supports non-thinking and thinking modes for code, knowledge, and agent tasks. The 2026-07-31 production update (DeepSeek-V4-Flash-0731) adds substantially stronger agent behavior and DSpark speculative decoding, and is included as a named checkpoint variant. The original Flash API has retired and its alias now routes to V4.1 Flash; original and updated V4 weights remain downloadable.

DeepSeek-V4-Pro

DeepSeek

DeepSeek-V4-Pro is the flagship tier of the V4 family, first publicly released as a preview on April 24, 2026. Its hybrid Compressed Sparse Attention and Heavily Compressed Attention architecture supports million-token contexts alongside manifold-constrained hyper-connections. It supports non-thinking and thinking modes for code, knowledge, and agent tasks. The 2026-08-13 production update (DeepSeek-V4-Pro-0813) adds substantially stronger agent behavior and DSpark speculative decoding, and is included as a named checkpoint variant. The Pro API remains available after the V4.1 Flash launch.

LLaDA2.0-Uni

InclusionAI (Ant Group)

LLaDA2.0-Uni unifies visual understanding, image generation, and instruction-based editing within a diffusion language model. Built from LLaDA2.0-Mini, its MoE backbone predicts masked semantic tokens, while SigLIP-VQ encodes images and a diffusion decoder reconstructs visual outputs. It supports visual question answering, image captioning, document understanding, and preliminary interleaved reasoning and generation. Apache-2.0 weights and code were released April 23, with FP8 and additional serving integrations arriving later. Its language-based visual understanding places it within the multimodal language-model scope.

GPT-5.5 Pro

OpenAI

GPT-5.5 Pro is the higher-compute GPT-5.5 tier for demanding research and professional questions. It launched in ChatGPT alongside GPT-5.5 before the API release. The tier aims for more accurate and comprehensive answers, using additional reasoning time and a higher token price. Its Responses API interface supports long-context, asynchronous work.

GPT-5.5

OpenAI

GPT-5.5 is a frontier general model for coding, knowledge work, computer use, and scientific analysis. It plans and checks multi-step tasks while moving across tools, with an emphasis on stronger results using fewer tokens. It launched in ChatGPT and Codex on April 23, with API access following April 24. The API context is larger than the context provided by Codex at launch.

Hy3 Preview

Tencent

Hy3 Preview is the first major model released on Tencent’s rebuilt Hunyuan training infrastructure. Its mixture-of-experts language backbone has 295B total parameters and 21B active per token, with additional multi-token-prediction parameters. The model supports 256K context and targets reasoning, coding, instruction following, and agent workflows. Public weights use the Tencent Hy Community License, which differs from the later Apache-licensed stable Hy3 release. Preview limitations and licensing should be evaluated before deployment.

Grok Voice Think Fast 1.0

xAI

Grok Voice Think Fast 1.0 is xAI’s reasoning-enabled spoken-dialogue model. It combines low-latency speech interaction with reasoning and tool use for voice assistants. The model can answer spoken questions and take multi-step actions through application tools, so it is distinct from speech transcription or standalone text-to-speech. The first version adds reasoning while maintaining conversational responsiveness. Its parameter count and detailed architecture are not disclosed. Real-time latency and interaction quality also depend on transport, audio capture, and the surrounding application.

MiMo V2.5 Pro

Xiaomi

MiMo V2.5 Pro is Xiaomi’s text-only flagship for complex coding and sustained agent workflows. It has approximately 1.02T total parameters and 42B active parameters, with hybrid sliding-window attention supporting 1M context. The public API beta preceded the April 27 publication of MIT-licensed base and post-trained weights. Pro and the native multimodal V2.5 tier are separate models. UltraSpeed and FP4 serving options retain this identity rather than creating additional model pages.

MiMo V2.5

Xiaomi

MiMo V2.5 is Xiaomi’s full-modal language model with text, image, audio, and video understanding. Its sparse backbone has approximately 310B parameters and activates 15B per token, supporting a 1M-token context. It targets general agents, cross-modal reasoning, and tool-driven workflows. Public API testing preceded the April 27 weight publication. Both base and post-trained weights are now MIT-licensed; quantized serving and draft models are grouped rather than treated as new identities.

Ling-2.6-Flash

InclusionAI (Ant Group)

Ling-2.6-Flash is Ant Group’s efficient instruction-following tier, designed for faster responses and lower token overhead in everyday agents. The 104B MoE network activates 7.4B parameters, combining MLA and Lightning Linear Attention in a 1:7 ratio. Post-training focuses on concise execution, coding, planning, and tool use rather than long visible reasoning. The first-party April 22 changelog documents public API and chat availability, preceding the later weight upload. MIT weights enable local serving, with a native 128K context and demonstrated YaRN extension to 256K.

BAR

Allen Institute for AI

BAR, or Branch-Adapt-Route, is Ai2’s released modular language-model family built from an Olmo 2 7B anchor and independently trained capability experts. Math, coding, tool-use, and safety experts are combined through a router into a final five-expert mixture, enabling capability updates without retraining the full model. The public checkpoints include the original anchor, intermediate two-expert combinations, and final combined model as variants. This is a concrete weight release supporting a post-training method, rather than a framework-only announcement. Expert merging and routing must be evaluated for interactions and task quality.

Kimi K2.6

Moonshot AI

Kimi K2.6 advances the Kimi multimodal agent family with stronger long-horizon coding, coding-driven design, and tool-based execution. It targets complicated end-to-end engineering tasks across languages such as Rust, Go, and Python, including performance optimization and DevOps. Native image and video understanding supports visual interfaces and design workflows. The model is available through Kimi chat, Kimi Code, the official API, and downloadable modified-MIT weights. Agent-swarm demonstrations use orchestration around the model; their concurrency should not be confused with its expert count.

Qwen3.6-Max-Preview

Alibaba

Qwen3.6-Max-Preview is a publicly accessible proprietary preview targeting stronger agentic coding, world knowledge, and instruction following than Qwen3.6-Plus. Qwen documents Qwen Studio access and a named Alibaba Cloud Model Studio endpoint, with preserved thinking recommended for agent workloads. It qualifies as a released preview even though development continued after launch. The announcement’s changing API rollout language means early preview and later hosted access should not be confused with a final GA model.

Grok 4.3

xAI

Grok 4.3 is a proprietary xAI language-model release. A configurable-reasoning model for coding, tool workflows, and general knowledge work. It accepts images and text, supports structured output and function calling, and permits none, low, medium, or high reasoning effort. The provider model documentation establishes the capabilities; the public integration catalog supplies the April release date because the current provider page does not state it. Weights are not publicly downloadable. Published agent and coding results depend on the evaluation harness, tool permissions, and inference settings, so no unsupported standalone benchmark score is assigned.

Qwen3.6-Flash

Alibaba

Qwen3.6-Flash is the efficiency-oriented hosted Qwen3.6 tier for multimodal reasoning, coding, and agent workflows. Official documentation lists a million-token context, text output, function calling, built-in tools, and structured output. It provides a lower-cost service option than Plus while accepting images and video. Its API identity should not be treated as interchangeable with a particular downloadable Qwen3.6 checkpoint, because hosted production features and context limits differ.

Qwen3.6

Alibaba

Qwen3.6 is an open-weight multimodal update focused on practical agentic coding and development stability. The first checkpoint is a 35B-A3B MoE model, followed on April 22 by a 27B dense model. Both build on the Qwen3.5 hybrid-attention architecture and support preserving reasoning history across turns. Images and video enable visually grounded coding and reasoning. Apache-2.0 weights support local deployment with a native 256K context; larger contexts require explicit extension. The proprietary hosted Plus and Flash tiers are distinct service identities.

Claude Opus 4.7

Anthropic

Claude Opus 4.7 advances Opus coding, visual understanding, and complex multi-step work. It is designed to check its own outputs more carefully and sustain difficult professional tasks. The release adds higher-resolution image handling and extra-high effort, with adaptive thinking and a million-token context. A changed tokenizer and reasoning behavior mean token consumption and prompts may differ from earlier Opus models.

GPT-Rosalind

OpenAI

GPT-Rosalind is OpenAI’s purpose-built reasoning model for life sciences research across chemistry, protein engineering, genomics, and translational medicine. It combines domain understanding with scientific tool orchestration for evidence synthesis, hypotheses, experimental planning, and data analysis. It launched as a trusted-access research preview and later became generally available to eligible organizations. Access through ChatGPT, Codex, and the API requires approval; it is a research assistant whose conclusions still require scientific validation.

SWE-check

Cognition

SWE-check is a specialized bug-detection language model developed by Cognition with Applied Compute and first made available in the Windsurf Next public preview. It analyzes a change diff, investigates the surrounding code, and produces structured bug descriptions and fixes for display in the editor. Reinforcement learning first builds task capability, then tunes latency and product behavior using a separate alignment phase. This is a distinct trained generative review model rather than a generic code-review feature or binary decision endpoint. Architecture, parameter counts, and model context limits remain undisclosed. Its reported parity with frontier models applies to internal in-distribution evaluations, with a remaining gap on out-of-distribution tasks.

Gemini Robotics ER 1.6

Google

Gemini Robotics ER 1.6 is the independently released embodied-reasoning update for interpreting physical spaces and planning robot tasks. It advances instrument reading and spatial and physical reasoning, with multimodal input and text output. Code execution, function calling, and structured responses let it coordinate external robot APIs. It was a public API preview, subsequently retired in August 2026.

Audio Flamingo Next

NVIDIA

Audio Flamingo Next is an open-weight NVIDIA audio-language research model. The successor to Audio Flamingo 3 improves speech, environmental-sound, and music reasoning and handles audio up to thirty minutes. Temporal Audio Chain-of-Thought grounds intermediate reasoning in timestamps. Instruct, thinking, and captioner checkpoints are variants of the same research release. The release supports research use under a non-commercial license. Descriptions of perceptual content are generated language and can misidentify events, speakers, or musical attributes.

EXAONE 4.5

LG AI Research

EXAONE 4.5 is LG AI Research’s first open-weight vision-language model, combining a 33B total parameter architecture with a 1.2B vision encoder. It extends the EXAONE family into image understanding, technical documents, charts, and multilingual visual reasoning, with a 256K context window. The release publishes deployment code and evaluation information alongside weights. Its custom EXAONE license must be consulted for usage rights rather than treating the model as Apache-licensed. It is suited to research and domain-specific document assistants, while the actual reliability of extracted facts and visual reasoning requires application-level validation.

LFM2.5 VL 450M

Liquid AI

The compact visual tier improves multilingual image understanding, grounding, and instruction following through extended pretraining and reinforcement learning. It adds bounding-box prediction and text function calling to visual question answering. Its small footprint supports real-time edge applications that turn image observations into structured language outputs. Downloadable weights use the Liquid Open Model License, whose commercial-use conditions should be checked before deployment. Local inference supports private customization, while task reliability and performance require evaluation on the intended hardware.

Muse Spark

Meta

Muse Spark is the first model developed by Meta Superintelligence Labs in the Muse family. It is a proprietary multimodal reasoning model with tool use, visual reasoning, and multi-agent orchestration, initially released through Meta AI on the web and in the app. It offers different response and reasoning modes for everyday questions and harder tasks. Its release marked a new training program rather than an open Llama checkpoint. Architecture and parameter count are not publicly disclosed, and planned future open-weight releases do not apply to this model.

Claude Mythos Preview

Anthropic

Claude Mythos Preview is Anthropic’s published frontier research model, documented in an April 2026 system card and deployed to selected Project Glasswing partners. The general-purpose model demonstrated substantial coding and vulnerability-research capabilities, motivating a restricted defensive-security program rather than general release. It is included as a substantive published research checkpoint, distinct from the later Fable/Mythos 5 family. Partner APIs existed, but no general public inference or open weights were offered.

Qwen3.6-Plus

Alibaba

Qwen3.6-Plus is a proprietary hosted multimodal model improving practical coding agents, repository-level reasoning, and visual understanding over Qwen3.5-Plus. The service supports a million-token context by default and preserves thinking across turns for multi-step tasks. Official APIs include function calling, built-in tools, and structured output. It is a distinct hosted tier from the subsequently released Apache-2.0 Qwen3.6 35B and 27B checkpoints.

Gemma 4

Google

Gemma 4 is an open-weight multimodal family with native thinking and function calling, released under Apache 2.0. The initial E2B and E4B models target constrained devices, while 26B A4B uses sparse experts and 31B is dense. All support vision and video frames; audio is available in the smaller models. Context lengths vary by size, and the independently released encoder-free 12B update has its own page.

GLM-5V-Turbo

Z.ai

GLM-5V-Turbo is Z.ai’s hosted multimodal coding foundation model. It processes text, images, videos, and rendered file content for visual coding, long-horizon planning, and tool-based action execution. Typical workflows include recreating interfaces from screenshots, inspecting rendered pages while debugging, and exploring websites through an agent. Official documentation lists a 200K context and 128K maximum output. A technical report describes the native multimodal model; no public downloadable checkpoint is established for this specific Turbo tier.

Trinity Large Thinking

Arcee AI

Trinity Large Thinking is Arcee AI’s reasoning-focused post-training of Trinity Large Base. The sparse mixture-of-experts model has 398B total parameters and approximately 13B active per token, with documented context extension to 512K. It generates explicit thinking traces and is trained for long-running agents, multi-turn tool calls, and coherent instruction following. Thinking tokens must remain in conversation history for supported agent behavior. API access and downloadable weights are available. The current model card specifies OpenMDW-1.1, despite Apache 2.0 wording in the original announcement; this record follows the current weight license.

Rnj 1.5 Instruct

Essential AI

Rnj 1.5 is Essential AI’s long-context instruction release, extending the Rnj family from 32K to 160K context and improving repository-level coding. Mid-training emphasizes long-document comprehension and repository fill-in-the-middle tasks. This is a materially updated checkpoint rather than a serving alias; primary artifact history establishes April 2026, while the exact public release day remains unverified. Apache-licensed weights support local deployment and customization. The model card warns of identity confusion and inaccurate factual recall; long generations can be less reliable than long-context comprehension.

Holo3 122B-A10B

H Company

Holo3 122B-A10B is H Company’s larger hosted computer-use model, positioned for longer and more complex enterprise workflows. Its mixture-of-experts architecture has 122B total language parameters with about 10B active per token, and combines screenshot understanding with multi-step navigation and structured actions. It was publicly introduced March 31, before the April 1 Hugging Face article. The current provider catalog documents a 65,536-token context, paid API access, and research-only non-commercial weight terms. An openly downloadable 122B checkpoint was not verified, so public access is recorded through the API rather than extrapolated from the Apache-licensed 35B tier.

Holo3 35B-A3B

H Company

Holo3 35B-A3B is H Company’s Apache-licensed sparse vision-language model for enterprise computer use. Based on Qwen3.5-35B-A3B, it is trained on open, synthetic, and human-annotated trajectories to improve perception, multi-step decisions, and navigation across web and desktop software. It was introduced on March 31 alongside a larger hosted Holo3 tier; the subsequent Hugging Face team article appeared April 1. This checkpoint provides downloadable weights and was originally served through the free API tier. Current API documentation instead lists Holo3.1 for the smaller tier, so continued first-party hosted inference on the original Holo3 35B is not asserted.

LFM2.5 350M

Liquid AI

This separately launched compact tier adds extended pretraining and reinforcement learning to Liquid’s 350M architecture. It focuses on structured output, extraction, and tool selection rather than lengthy general reasoning. Both base and post-trained weights are available for developers who need low-cost specialized inference on small edge devices. Downloadable weights use the Liquid Open Model License, whose commercial-use conditions should be checked before deployment. Local inference supports private customization, while task reliability and performance require evaluation on the intended hardware.

Bonsai

PrismML

Bonsai is PrismML’s independently published low-bit language-model family for local and on-device deployment. The original release contains 8B, 4B, and 1.7B models using binary weights across embeddings, attention, feed-forward layers, and the language head. The representative 8.19B checkpoint uses a Qwen3-8B-derived dense Transformer, supports 65,536 tokens, and occupies approximately 1.15 GB for weights. April’s Ternary Bonsai variants offer a different accuracy and memory tradeoff and are grouped here alongside routine packings. Apache-licensed weights and specialized inference kernels are available. This family page excludes the later, separately positioned 27B vision-language tier.

Mercury Edit 2

Inception

Mercury Edit 2 is Inception’s specialized diffusion model for predicting a developer’s next code edit from recent changes and repository context. Preference training targets useful, selective suggestions rather than overly frequent edits. It complements completion endpoints and is available through the Inception API and editor integrations. Public weights and parameter counts are undisclosed. Reported latency depends on hardware, concurrency, and prompt distribution; production applications require workload-specific evaluation.

GLM-5.1

Z.ai

GLM-5.1 strengthens the GLM engineering model’s ability to stay productive over extended agent sessions. Post-training improves problem decomposition, experiment execution, tool use, and strategy revision across hundreds of rounds. It targets repository generation, complex debugging, and long-horizon software optimization. MIT weights and an official API permit hosted and local use. This page dates the verified public model release to the official April 7 announcement; earlier subscription rollouts are not established by the cited release record. The March 27 official Coding Plan announcement establishes public subscription access before the April 7 broader release.

Gemini 3.1 Flash Live

Google

Gemini 3.1 Flash Live is the public-preview Gemini 3.1 audio-to-audio model for real-time conversational agents. It accepts text, images, audio, and video and returns text and audio, with reasoning and tools during live sessions. It supports native audio output and search grounding. The current catalog describes it as a legacy preview and recommends migration to Gemini 3.8 Live.

MolmoWeb

Allen Institute for AI

MolmoWeb is a released vision-language model trained for browser tasks from screenshots and natural-language instructions. Its 4B and 8B weights are distinct from the browser harness: the checkpoints generate textual reasoning, browser actions, and task-completion responses. Training uses MolmoWebMix, an open collection of synthetic and human browsing trajectories. Native and Transformers-compatible checkpoints are serving variants; compatibility fixes after launch do not establish a new model generation. Completing a web task depends on the external browser, observation/action interface, and execution permissions.

Composer 2

Cursor

Composer 2 is Cursor’s Kimi K2.5-derived agentic coding model, released after continued pretraining on code-focused data and large-scale reinforcement learning in realistic Cursor sessions. The training combines stronger base knowledge with the same tools and agent harness used for deployment. It supports multi-step repository work and long-running coding sessions, with a documented 200K single-call context and a faster serving option within the same model identity. The launch on March 19 predates the public technical report. Its ancestry is explicitly documented; this does not make the proprietary post-trained weights publicly downloadable. Performance and cost claims depend on task sets, harnesses, and serving configuration.

Nemotron-Cascade 2

NVIDIA

Nemotron-Cascade 2 is a sparse reasoning model post-trained through cascade reinforcement learning and multi-domain on-policy distillation. The training recipe transfers specialist strengths across mathematical reasoning, coding, science, and instruction following into a compact active-parameter model. It supports thinking and non-thinking modes and very long context. NVIDIA releases weights, training datasets, and reproducibility materials, making the checkpoint useful for reasoning research as well as self-hosted assistants. Competition-level results may rely on extended inference or tool setups.

Qwen3.5-Max-Preview

Alibaba

Qwen3.5-Max-Preview is a publicly accessible Qwen model preview deployed on Arena for community evaluation. The official announcement invites users to test its capabilities while final optimization continues. This page represents the actual Arena preview identity; it does not assert that a final Qwen3.5-Max checkpoint, API, or downloadable weights were released. The public announcement does not disclose complete architecture or parameter specifications.

MolmoPoint

Allen Institute for AI

MolmoPoint is Ai2’s vision-language release with a token-based pointing mechanism that selects image regions directly from visual features rather than spelling out coordinates as ordinary text. It combines natural-language image and video understanding with spatial grounding, video pointing, and tracking. The 8B model uses Qwen3-8B and SigLIP; related 4B video and 8B GUI checkpoints specialize the same architecture. Weights, training code, and data support research on reliable visual interaction. Available implementations differ in which grounding functions they expose, and visual selections can still be inaccurate.

MiniMax M2.7

MiniMax

MiniMax M2.7 improves professional software engineering, office work, and character consistency, and explores model participation in its own training and harness improvement. It is a text MoE model in the M2 family, exposed through MiniMax API and Agent products. Downloadable weights were released after the initial hosted launch. The current weight license permits non-commercial use; commercial weight deployment requires prior written authorization. Agent Team workflows are orchestration around the model, not additional model modalities.

MiMo V2 Omni

Xiaomi

MiMo V2 Omni is Xiaomi’s native multimodal model for visual, audio, video, and text understanding. It combines modality encoders with a shared language backbone so an agent can ground its responses in screenshots, recordings, and other inputs. Hosted inference supports reasoning and structured tool use. It is a distinct multimodal tier of the V2 generation, with parameter counts left unspecified because the launch materials do not disclose a reliable total. Its availability does not imply downloadable weights.

MiMo V2 Pro

Xiaomi

MiMo V2 Pro is Xiaomi’s hosted flagship for long-running software engineering and agent tasks. Its trillion-scale sparse backbone activates approximately 42B parameters and combines sliding-window and global attention. The text model supports up to 1M tokens and structured tool use. Anonymous evaluation names such as Hunter Alpha are aliases, rather than separate releases. Xiaomi provides API and chat access; the V2 Pro weights were not publicly released alongside the later V2.5 family.

GPT-5.4 Mini

OpenAI

GPT-5.4 Mini brings the GPT-5.4 generation to a faster, lower-cost tier for repeated coding and agent tasks. It supports text and images, reasoning, structured outputs, compaction, tool search, and built-in computer use. Its API context is smaller than the flagship GPT-5.4 window, making it a practical worker or subagent when a task fits within that limit.

GPT-5.4 nano

OpenAI

GPT-5.4 Nano is the least expensive GPT-5.4-class tier, focused on simple high-volume operations. It accepts text and images, supports configurable reasoning and structured output, and provides compaction. Unlike Mini, it does not support built-in computer use or tool search. Its intended uses include classification, extraction, ranking, and focused subagent work.

Holotron-12B

H Company

Holotron-12B is H Company’s first Holotron computer-use model, developed with NVIDIA and post-trained from Nemotron Nano 2 VL. The approximately 12B hybrid state-space/attention vision-language model is optimized for production throughput, long interactive histories, and multiple screenshots. Targeted training turns the general multimodal backbone into an agent policy that interprets screens and emits actions across web, desktop, and mobile environments. The released model card explicitly dates publication to March 16 and identifies NVIDIA model licensing. Public weights allow local experimentation; an external harness supplies screenshots, executes actions, and manages the agent loop. Benchmark throughput is specific to the tested GPU and workload configuration.

Leanstral

Mistral AI

Leanstral is Mistral’s open-weight specialist model for Lean 4 formal mathematics and program verification. The first Leanstral release focuses on realistic formal repositories and uses Lean compiler feedback as a correctness check. It edits proof files, runs tools, and revises proofs after compiler feedback within an agent scaffold. Apache-2.0 weights permit local deployment and adaptation. Published success rates involve explicit inference budgets and multiple attempts; a long agent run is not the same as a single context window.

Mistral Small 4

Mistral AI

Mistral Small 4 unifies chat instruction following, reasoning, coding, and image understanding in one configurable open-weight model. Its sparse expert architecture uses 128 experts, selecting four per token; the provider distinguishes six billion active expert parameters from eight billion including embedding and output layers. Users can choose direct answers or deeper reasoning per request. It is distributed under Apache 2.0 with support in common self-hosting runtimes. Despite the Small name, its total weights require considerably more memory than the earlier 24B Small models.

Reka Edge

Reka AI

Reka Edge is a 7B vision-language model for image and video understanding, physical grounding, and language-guided tool use. It combines a 657M ConvNeXt V2 visual encoder with a 6.4B language transformer trained from scratch. Compact visual tokens improve streaming efficiency. Hosted API and demo access complement downloadable weights. The release license permits commercial use below a revenue threshold and imposes additional conditions otherwise; it is not equivalent to an unrestricted Apache release. EdgeQ deployment is grouped as a quantized variant.

Nemotron 3 Super

NVIDIA

Nemotron 3 Super is an open-weight NVIDIA language-model release. A long-context reasoning model for software engineering and multi-step agents. It uses latent-space routed experts, selective attention, multi-token prediction, and an NVFP4 training recipe. The public research release is dated March 10; the weight card records March 11 distribution. The license permits commercial and non-commercial use subject to its published terms.

Grok 4.20

xAI

Grok 4.20 is a proprietary xAI language-model release. The publicly documented API release improves coding, knowledge work, instruction following, and hallucination behavior. Reasoning and non-reasoning endpoints are joined by a multi-agent serving option that runs parallel cooperative agents. The multi-agent orchestration is grouped as a variant rather than assigned a speculative new architecture. Earlier consumer beta reports do not establish a precise primary-source availability date. Weights are not publicly downloadable. Published agent and coding results depend on the evaluation harness, tool permissions, and inference settings, so no unsupported standalone benchmark score is assigned.

Sarvam 30B

Sarvam AI

Sarvam 30B is an Indian multilingual language model trained from scratch for practical local deployment. Its sparse experts support coding, reasoning, tool use, and conversational quality across 22 Indian languages. The 30B tier uses fewer layers and grouped key-value attention to reduce inference overhead. Public weights were released on March 6, following the family’s February announcement. The checkpoint supports up to 128K context, while reported evaluations use shorter task-specific limits. Model outputs require task-specific checks in languages and domains beyond the training focus.

Olmo Hybrid

Allen Institute for AI

Olmo Hybrid is a fully open 7B language model that combines Transformer attention with gated DeltaNet linear recurrent layers. Three quarters of its sequence-mixing layers use gated DeltaNet, reducing long-context state costs while retaining attention in the remaining layers. Ai2 releases base, instruction, and thinking checkpoints alongside data and training artifacts. The model studies scaling and expressivity compared with Olmo 3 under controlled training, and provides a practical platform for hybrid-model experiments. Throughput comparisons depend on sequence length and hardware rather than representing a fixed speedup for every task.

KARL

Databricks

KARL is Databricks AI Research’s published language-model specialization for grounded enterprise search and knowledge reasoning. The technical report explicitly defines KARL as GLM 4.5 Air post-trained with two iterations of multi-task reinforcement learning, rather than only the surrounding search-agent harness. Synthetic question generation, long-horizon solving, and large-batch off-policy learning train retrieval, cross-document synthesis, and numerical reasoning behaviors. Parallel rollouts and aggregation provide optional test-time scaling. The research model is documented in a public report, but downloadable weights and a named public inference endpoint are not established; platform product access does not prove an identical checkpoint.

GPT-5.4 Pro

OpenAI

GPT-5.4 Pro is the higher-compute GPT-5.4 tier for difficult professional and technical questions. It shares the documented text and image interface and expanded API context but spends more reasoning computation for stronger answers. Responses API background execution accommodates long-running requests. It is a distinct API tier rather than a dated snapshot of GPT-5.4.

GPT-5.4

OpenAI

GPT-5.4 is a general model for professional work that brings reasoning, coding, and native computer interaction into one release. The Responses API supports computer use, tool search, and context compaction for extended agent workflows, alongside text and image input. Its expanded API context can hold large collections of documents, though platform limits and long-context pricing differ.

Phi-4-reasoning-vision-15B

Microsoft

Phi-4-reasoning-vision-15B is an MIT-licensed open-weight model in Microsoft’s Phi family. It adds visual perception to the Phi-4-reasoning backbone, supporting visual math, chart interpretation, screenshots, and image-conditioned reasoning. Weights and sample inference code are available on Hugging Face, with deployment through supported local runtimes or hosted platforms. This checkpoint is a static trained model; accuracy and language coverage vary by task. Small model efficiency should not be taken as equal capability to much larger generalist systems.

Gemini 3.1 Flash-Lite

Google

Gemini 3.1 Flash-Lite is the lower-cost Gemini 3.1 tier for high-frequency, lightweight workloads. Its public preview began in March, followed by general availability in May. It accepts text, image, audio, and video inputs and supports thinking, function calling, and structured responses. It is intended for repeated extraction, summarization, translation, and classification jobs with precise instructions.

GPT-5.3 Instant

OpenAI

GPT-5.3 Instant is the conversational GPT-5.3 model used in ChatGPT and exposed through the gpt-5.3-chat-latest alias. It prioritizes responsive everyday conversation and image understanding rather than long reasoning runs. The alias can change snapshots and is not a fixed checkpoint. This record preserves the distinct released Instant model despite subsequent deprecation.

SWE-1.6

Cognition

SWE-1.6 is Cognition’s next coding model, first served as a limited-user preview on March 1 and generally released in Windsurf on April 7, 2026. It uses the same underlying pretrained model as SWE-1.5, but the production release was post-trained anew to improve both coding capability and interaction behavior. Training discourages overthinking, repeated reasoning loops, and unnecessary terminal use, while encouraging parallel tool calls. The earlier preview and production checkpoint are grouped under this identity, with their behavioral differences made explicit. Standard and faster serving options provide the same model intelligence; neither establishes publicly downloadable weights or a standalone model API.

Qwen3.5-Omni-Flash

Alibaba

Qwen3.5-Omni-Flash is a proprietary native omnimodal model for text, image, audio, and video understanding with text and speech output. The Thinker–Talker architecture combines hybrid reasoning, audio-visual comprehension, and streaming speech synthesis, targeting multimodal assistants and real-time interaction. This separately named hosted tier is available through Qwen APIs, including a Realtime interface. The family was publicly released in March 2026; the dated March 15 serving snapshot is not used as an unverified launch date. No public downloadable weights are established for this tier.

Qwen3.5-Omni-Plus

Alibaba

Qwen3.5-Omni-Plus is a proprietary native omnimodal model for text, image, audio, and video understanding with text and speech output. The Thinker–Talker architecture combines hybrid reasoning, audio-visual comprehension, and streaming speech synthesis, targeting multimodal assistants and real-time interaction. This separately named hosted tier is available through Qwen APIs, including a Realtime interface. The family was publicly released in March 2026; the dated March 15 serving snapshot is not used as an unverified launch date. No public downloadable weights are established for this tier.

GLM-5-Turbo

Z.ai

GLM-5-Turbo is a hosted language model optimized during training for OpenClaw agent workflows. It strengthens tool invocation, instruction decomposition, scheduled and persistent tasks, and long-chain execution. A roughly 200K context and up to 128K output support multi-step agents, with function calling, JSON structured outputs, context caching, and MCP integration. Official documentation provides a named API endpoint. The weights and detailed architecture are not publicly established; it should not inherit open-weight status from the standard GLM-5.

Mercury 2

Inception

Mercury 2 is Inception’s hosted reasoning language model based on diffusion. Parallel refinement generates multiple tokens per iteration instead of purely sequential decoding. Tunable reasoning, native tool use, and schema-aligned JSON support coding, retrieval pipelines, and interactive agent loops. Public weights and parameter counts are undisclosed. Reported latency depends on hardware, concurrency, and prompt distribution; production applications require workload-specific evaluation.

LFM2 24B A2B

Liquid AI

Liquid’s largest second-generation language model combines sparse experts, short convolutions, and grouped-query attention. It targets tool-using agents that can run on consumer computers while retaining cloud-serving options. The checkpoint card documents a 32K trained context; larger configuration ceilings are not treated as validated operating limits. Downloadable weights use the Liquid Open Model License, whose commercial-use conditions should be checked before deployment. Local inference supports private customization, while task reliability and performance require evaluation on the intended hardware.

Qwen3.5-Flash

Alibaba

Qwen3.5-Flash is the lower-cost hosted tier of Qwen3.5, supporting text, image, and video input with text output. A million-token context and built-in tools support long-document analysis, coding agents, and visually grounded workflows. Official documentation lists function calling and structured output alongside dated API snapshots. Its serving architecture and parameter count are not sufficiently disclosed to borrow specifications from a particular open Qwen3.5 size.

GPT-Audio-1.5

OpenAI

GPT-Audio-1.5 is OpenAI’s audio conversation model for speech and text inputs and outputs through the Chat Completions API. It is designed for language interaction rather than transcription-only or TTS-only use. The 1.5 release updates the conversational audio model, retaining the documented endpoint and language interface.

GPT-Realtime-1.5

OpenAI

GPT-Realtime-1.5 is OpenAI’s realtime conversational model for low-latency speech and text interaction. Its bidirectional Realtime API supports voice-agent sessions, image input, and external tools. The 1.5 release updates the conversational audio model, retaining the documented endpoint and language interface.

Sarvam 105B

Sarvam AI

Sarvam 105B is a multilingual sparse language model trained from scratch in India. It emphasizes reasoning, coding, and agents across 22 Indian languages, using an MLA-style attention stack and context extension. Public access began through the limited Indus beta on February 20; Apache-licensed weight publication followed on March 6. This page represents the original open checkpoint, not later hosted conversational and work-agent specializations. The checkpoint supports up to 128K context, while reported evaluations use shorter task-specific limits. Model outputs require task-specific checks in languages and domains beyond the training focus.

Gemini 3.1 Pro

Google

Gemini 3.1 Pro is an independently released Pro update for complex reasoning, software engineering, and multimodal work. It is available as a public API preview with a million-token input context and adjustable thinking. A separate custom-tools endpoint prioritizes developer-supplied tools when built-in and custom tools are combined; that serving variant is represented on this page.

Claude Sonnet 4.6

Anthropic

Claude Sonnet 4.6 is the independently released Sonnet update for coding, computer use, long-context reasoning, and professional work. It improved follow-through, instruction following, and visual design and launched with a million-token context in beta. The current API documentation lists that context as standard. It offers adaptive thinking and tools through the Claude API and cloud partners, with text and image inputs and text output.

Tiny Aya

Cohere

Tiny Aya is Cohere Labs’ 3.35B multilingual model family designed to make local language generation practical across 70 languages. The family includes a base checkpoint, a globally balanced instruction model, and Earth, Fire, and Water regional variants. These emphasize African and West Asian, South Asian, and European and Asia-Pacific languages respectively. Instruction checkpoints are available through Cohere’s Chat API and downloadable under a non-commercial license. The family’s 8K context and compact size suit offline assistants, translation experiments, and regional-language applications without requiring a large hosted model.

Qwen3.5-Plus

Alibaba

Qwen3.5-Plus is the hosted production tier corresponding to Qwen3.5-397B-A17B. It adds a million-token context by default, official built-in tools, and adaptive tool use to the multimodal foundation model. It supports reasoning, coding, image/video understanding, function calling, and structured output. Hosted model terms govern service access; the related Apache-2.0 checkpoint has its own open-family page and does not establish downloadable weights for every proprietary serving update.

Qwen3.5

Alibaba

Qwen3.5 is a native vision-language foundation-model family combining Gated DeltaNet and gated attention. The initial 397B-A17B MoE model was followed by 122B-A10B, 35B-A3B, and dense sizes from 27B to 0.8B. Early fusion of text and visual data supports images, video, coding, reasoning, and tool-based agents, with expanded coverage of 201 languages and dialects. Apache-2.0 weights permit local use. This page groups the open sizes; separately positioned Plus and Flash hosted tiers have different service limits and their own pages.

WebWorld

Alibaba

WebWorld is Qwen’s open language-world-model family for browser-agent training and evaluation. It predicts environment states in accessibility-tree, HTML, XML, Markdown, and natural-language formats after a sequence of actions. Models built from Qwen3-8B, 14B, and 32B are trained on more than a million real web trajectories using a curriculum that first learns dynamics and then activates causal reasoning. The February16 technical report establishes the published research identity. Apache-2.0 weights and data support simulated long-horizon interaction; this page groups all three routine sizes.

Ling-2.5-1T

InclusionAI (Ant Group)

Ling-2.5-1T updates Ant Group’s trillion-parameter language family with hybrid linear attention for efficient long contexts. It combines Multi-head Latent Attention and Lightning Linear Attention in a 1:7 ratio with sparse experts. This instant-model tier targets concise reasoning, preference alignment, creative writing, and native agent interaction, with a native 256K context extendable to one million tokens. Both downloadable MIT weights and historical hosted access are documented. The first-party 2.5 APIs were retired September 30; this page preserves the released model identity.

Ring-2.5-1T

InclusionAI (Ant Group)

Ring-2.5-1T updates Ant Group’s trillion-parameter language family with hybrid linear attention for efficient long contexts. It combines Multi-head Latent Attention and Lightning Linear Attention in a 1:7 ratio with sparse experts. This thinking tier strengthens deep reasoning and sustained agent execution through fully asynchronous agentic reinforcement learning, with a native 128K context extendable to 256K. Both downloadable MIT weights and historical hosted access are documented. The first-party 2.5 APIs were retired September 30; this page preserves the released model identity.

Doubao Seed 2.0 Code

ByteDance

Doubao Seed 2.0 Code is ByteDance’s dedicated software engineering model in the Seed 2.0 family. The release strengthens repository changes, debugging, and code delivery through an external agent environment. Public inference is offered through Volcengine and ByteDance applications, with no downloadable weights. The tier is a separate product identity; API snapshots and later service migrations are variants. Parameter counts and unsupported endpoint limits are left unspecified.

Doubao Seed 2.0 Lite

ByteDance

Doubao Seed 2.0 Lite is ByteDance’s efficient productivity tier in the Seed 2.0 family. It combines language reasoning with image and video understanding for document-heavy tasks, visual analysis, and multi-step agents. Public inference is offered through Volcengine and ByteDance applications, with no downloadable weights. The tier is a separate product identity; API snapshots and later service migrations are variants. Parameter counts and unsupported endpoint limits are left unspecified.

Doubao Seed 2.0 Mini

ByteDance

Doubao Seed 2.0 Mini is ByteDance’s efficient productivity tier in the Seed 2.0 family. It combines language reasoning with image and video understanding for document-heavy tasks, visual analysis, and multi-step agents. Public inference is offered through Volcengine and ByteDance applications, with no downloadable weights. The tier is a separate product identity; API snapshots and later service migrations are variants. Parameter counts and unsupported endpoint limits are left unspecified.

Doubao Seed 2.0 Pro

ByteDance

Doubao Seed 2.0 Pro is ByteDance’s flagship productivity model in the Seed 2.0 family. It combines language reasoning with image and video understanding for document-heavy tasks, visual analysis, and multi-step agents. Public inference is offered through Volcengine and ByteDance applications, with no downloadable weights. The tier is a separate product identity; API snapshots and later service migrations are variants. Parameter counts and unsupported endpoint limits are left unspecified.

MiniMax M2.5

MiniMax

MiniMax M2.5 extends the M2 family through reinforcement learning in complex real-world environments. It targets software engineering, search and tool use, and office work, with improved planning before code generation and stronger professional web research. The release provides public modified-MIT weights and official API/Agent access. A high-speed hosted service variant shares the model identity and is grouped here. Reported research results using context management depend on that agent implementation.

GPT-5.3-Codex-Spark

OpenAI

GPT-5.3-Codex-Spark is the smaller coding model designed for near-instant interactive development. Its public research preview runs on Cerebras hardware through Codex for ChatGPT Pro subscribers, with separate usage limits. At launch it supported a 128K context and text-only input. It favors quick, targeted edits and does not automatically run tests unless asked, making it complementary to longer-horizon coding models.

Ming-Flash-Omni 2.0

InclusionAI (Ant Group)

Ming-Flash-Omni2.0 is Ant Group’s language-capable any-to-any multimodal model, combining text, image, audio, and video understanding with text, image, and audio generation. Its Ling2.0-based MoE backbone has 100B total and 6B active parameters, with modality-specific encoders and generation components. The February 11 official release improves visual encyclopedic knowledge, natural speech, and image synthesis and editing. MIT weights and inference code are downloadable. An earlier Flash-Omni preview is grouped as a historical variant, while distinct Lite models have different sizes and specifications.

MiniCPM-SALA

OpenBMB

MiniCPM-SALA is OpenBMB’s 9.48B language model combining sparse and linear attention for million-token contexts. One quarter of layers use InfLLM-v2 sparse attention and three quarters use Lightning Attention, with Hybrid Positional Embedding supporting length generalization beyond short training sequences. Apache-licensed weights and inference examples target long-context local language tasks with reduced cache overhead. The official model card documents deployment at up to 1M tokens, extending beyond its static 512K configuration value. Speedup claims depend on hardware, sequence length, and implementation rather than applying uniformly to every prompt.

GLM-5

Z.ai

GLM-5 scales the GLM family to 744B parameters and introduces DeepSeek Sparse Attention for efficient long-context inference. It targets complex systems engineering and long-horizon agents, with expanded pretraining and asynchronous reinforcement learning using the slime infrastructure. The model accepts and generates text, including structured tool calls. MIT weights support self-hosting, and official API/chat products provide hosted access. The February 11 public weight release precedes the February 12 dated product blog and documentation announcement.

Arctic AWM 4B

Snowflake

Arctic AWM 4B is Snowflake AI Research and UNC-Chapel Hill’s language model for multi-turn tool use, trained from Qwen3 4B using agentic reinforcement learning in executable synthetic environments. It generates language and tool calls against environments exposed through MCP interfaces. The February 10 research publication documents the training approach, while the released Apache-licensed checkpoint provides a concrete agent model rather than only the environment-generation software. External tools, state transitions, and evaluation scaffolding are separate from the model weights.

Composer 1.5

Cursor

Composer 1.5 is Cursor’s significant coding-model update, built by expanding reinforcement learning approximately twentyfold on the same pretrained base used for Composer 1. It adds adaptive thinking and trained self-summarization, balancing fast responses to simple requests with longer reasoning on difficult repository tasks. When a session approaches its context limit, the model writes a summary and continues from that state; this extends task duration rather than the raw context capacity of one inference call. The proprietary mixture-of-experts model is delivered through Cursor. The publisher does not identify the pretrained checkpoint or parameter count, so this record does not invent an ancestry relationship to Composer 1’s post-trained weights.

LLaDA2.1

InclusionAI (Ant Group)

LLaDA2.1 is Ant Group’s MoE diffusion-language model family, generating text through block-parallel refinement rather than exclusively left-to-right decoding. Token editing adds error-correcting generation and Speed/Quality modes, balancing parallel decoding speed against answer quality in 100B Flash and 16B Mini checkpoints. Public Apache-2.0 weights and custom inference code support research and self-hosting. Performance and latency depend on diffusion settings and the serving implementation.

UI-Venus-1.5

InclusionAI (Ant Group)

UI-Venus-1.5 is a distinct vision-language model specialization for GUI interaction, combining screenshot understanding with natural-language task reasoning and structured actions. The 1.5 release offers 2B, 8B, and 30B-A3B checkpoints for grounding and end-to-end mobile agents, with Apache-2.0 weights. Local inference requires an external action parser and execution framework; the model itself emits language reasoning and actions.

Claude Opus 4.6

Anthropic

Claude Opus 4.6 is the independently released Opus update for sustained coding, reasoning, and professional agent tasks. It improves planning, code review, debugging, and work in large repositories. Adaptive thinking and effort settings control computation, while context compaction supports longer workflows. The million-token context launched in beta and is now listed as standard in current API documentation, with 128K normal output and an optional 300K batch beta. The model accepts text and images and generates text.