Explore The LLM Database

MiniMax M2.1

MiniMax

Enhanced version of MiniMax-M2 with significantly improved multi-language programming capabilities, built for real-world complex coding tasks. Features exceptional performance in WebDev, AppDev (Android/iOS native development), and office automation scenarios. Delivers more concise and efficient responses compared to M2 while maintaining powerful general intelligence. Demonstrates excellent generalization across various programming tools and Agent frameworks.

GLM-4.7

Z.ai

Advanced coding-focused language model with significant improvements in multilingual agentic coding, terminal-based tasks, UI generation, tool using, and complex reasoning. Features Interleaved Thinking, Preserved Thinking, and Turn-level Thinking capabilities for enhanced stability in complex multi-turn tasks. Achieves substantial gains over GLM-4.6 across coding benchmarks, web browsing, and mathematical reasoning.

FunctionGemma

Google

FunctionGemma is a specialized version of Gemma 3 270M fine-tuned for function calling and designed to run on edge devices. It bridges natural language and software execution, translating user commands into executable API actions. The model excels at unified action and chat capabilities, switching seamlessly between generating structured function calls and conversational responses. Built specifically for customization through fine-tuning, it demonstrated 85% accuracy on Mobile Actions after training (up from 58% baseline). Small enough to run on mobile phones and edge devices like NVIDIA Jetson Nano, it uses Gemma's 256k vocabulary to efficiently tokenize JSON and multilingual inputs.

Nemotron 3 Nano

NVIDIA

Nemotron 3 Nano is a 31.6B parameter hybrid large language model developed by NVIDIA, released on December 15, 2025. It employs a novel Hybrid Mamba-Transformer Mixture-of-Experts (MoE) architecture, activating only ~3.2B parameters per token for efficient inference. Optimized for agentic AI, reasoning, and coding, it supports a 1 million token context window and features a 'Reasoning ON/OFF' toggle with a configurable thinking budget. NVIDIA has released the model with open weights, training data, and training recipes under the NVIDIA Open Model License.

Nemotron 3 Super

NVIDIA

Nemotron 3 Super is a high-accuracy reasoning model designed for multi-agent systems, featuring a Hybrid Latent Mixture-of-Experts (MoE) Mamba-Transformer architecture. Announced by NVIDIA in December 2025 with availability expected in H1 2026, it is optimized for collaborative agent workflows and high-volume tasks like IT ticket automation. The model utilizes approximately 100 billion total parameters with 10 billion active per token, leveraging NVIDIA's NVFP4 training format and Multi-Token Prediction (MTP) for efficiency. It supports a 1 million token context window and is released under the NVIDIA Open Model License.

Nemotron 3 Ultra

NVIDIA

Nemotron 3 Ultra is the flagship large language model in NVIDIA's Nemotron 3 family, announced on December 15, 2025, with general availability expected in the first half of 2026. It features a massive 500 billion total parameters (with approximately 50 billion active parameters per token) and utilizes a novel Hybrid Mamba-Transformer Latent Mixture-of-Experts (MoE) architecture. This architecture, combining Mamba-2 state-space models with Transformer attention layers and Latent MoE routing, is designed for high efficiency and performance in complex, long-horizon tasks. With a 1 million token context window, Nemotron 3 Ultra is optimized for demanding enterprise AI applications such as deep research, strategic planning, and large-scale multi-agent coordination. It was trained using NVFP4 precision on NVIDIA Blackwell GPUs on a 3 trillion token dataset and incorporates Reinforcement Learning for agentic tasks. The model supports advanced reasoning, tool use, and is multilingual, with open weights released under the NVIDIA Open Model License.

GPT-5.2

OpenAI

GPT-5.2 is OpenAI's flagship frontier model series designed for professional knowledge work, advanced coding, and agentic workflows. Released in December 2025 as a response to competitive pressures, it features a massive 400,000-token context window and a 128,000-token maximum output capacity. The model utilizes a Mixture-of-Experts (MoE) architecture to balance inference efficiency with deep reasoning capabilities. It is available in three variants—Instant, Thinking, and Pro—each optimized for different points on the latency-intelligence curve. GPT-5.2 demonstrates state-of-the-art performance in tool calling reliability (98.7%), coding (SWE-Bench Verified 80.0%), and long-context retrieval.

Devstral 2

Mistral AI

Devstral 2 is a state-of-the-art, open-weight agentic coding model developed by Mistral AI. Released in December 2025, it features a dense Transformer architecture with 123 billion parameters and a 256k token context window. Designed specifically for software engineering tasks, Devstral 2 excels at codebase exploration, multi-file editing, and complex instruction following. It supports multimodal inputs (text, code, and images) and is optimized for agentic workflows, achieving a score of 72.2% on the SWE-bench Verified benchmark. The model is released alongside the Mistral Vibe CLI for terminal-based code automation. While the weights are publicly available, the 123B model is distributed under a Modified MIT license with commercial restrictions for large entities.

Claude 4.5 Opus

Anthropic

Claude 4.5 Opus is the most intelligent and capable model in the Claude 4.5 family, designed for the most demanding reasoning, research, and coding tasks. It introduces a unique 'effort' parameter (low, medium, high) that allows users to control the depth of the model's reasoning process and token usage. Opus 4.5 excels in handling ambiguity, complex problem-solving, and deep analysis, often succeeding where other models fail on multi-system bugs or intricate research questions. It is priced to be more accessible than previous Opus generations while delivering state-of-the-art performance.

Gemini 3 Pro

Google

Gemini 3 Pro is Google's flagship multimodal foundation model, released in November 2025. Built on a sparse Mixture-of-Experts (MoE) Transformer architecture, it features a 1 million token context window and native understanding of text, images, audio, and video. The model introduces 'Deep Think' capabilities for enhanced reasoning, controlled via a 'thinking_level' parameter, and is optimized for 'agentic' workflows and 'vibe coding'—the generation of full applications from natural language. It supports advanced function calling and tool use, making it suitable for complex software engineering and long-context analysis tasks.

GPT-5.1

OpenAI

GPT-5.1 is a frontier-grade multimodal language model family released by OpenAI in November 2025. It introduces a unified system architecture featuring a 'Smart Router' that dynamically allocates compute resources between two primary modes: 'Instant' (optimized for low latency and conversational warmth) and 'Thinking' (optimized for deep, adaptive reasoning). The model utilizes a Sparse Mixture-of-Experts (MoE) architecture with a central language backbone and attachable modules, allowing it to process text, audio, image, and video inputs natively. Key capabilities include adaptive test-time compute, where the model adjusts its reasoning depth based on query complexity, and enhanced personalization options with distinct personality presets. It demonstrates significant improvements in instruction following, coding (via the Codex variants), and mathematical reasoning compared to its predecessor, GPT-5.

Qwen3-VL 235B (MoE)

Alibaba

The multimodal flagship of the Qwen 3 series. It integrates 'Interleaved-MRoPE' for superior spatial-temporal modeling and 'DeepStack' for vision-language alignment. It supports native interleaved contexts of up to 256K tokens.

Claude 4.5 Haiku

Anthropic

Claude 4.5 Haiku is the fastest and most cost-efficient model in the Claude 4.5 family, designed to deliver near-frontier intelligence with low latency. It is optimized for high-volume tasks, real-time responses, and sub-agent orchestration. Despite its smaller size, it supports advanced features like 'extended thinking' and computer use, making it suitable for parallelized workflows where speed is critical. It is positioned as a drop-in replacement for previous mid-tier models with significantly better economics.

Claude 4.5 Sonnet

Anthropic

Claude 4.5 Sonnet is a high-intelligence frontier model designed for complex agentic workflows, coding, and computer use. It balances capability with speed and cost, positioning it as the workhorse of the Claude 4.5 family. It features 'extended thinking' capabilities (beta) and is optimized for long-horizon tasks, capable of maintaining focus for extended periods (up to 30+ hours in internal tests). It introduces significant improvements in tool use, particularly in interacting with software interfaces and executing complex multi-step plans.

Qwen3-Omni 30B (MoE)

Alibaba

A native end-to-end omni-modal model capable of processing text, image, audio, and video. It supports audio understanding of up to 40 minutes and features a 'Thinker-Talker' MoE architecture for low-latency streaming.

GPT-OSS-120B

OpenAI

GPT-OSS-120B is a state-of-the-art open-weight language model that delivers strong real-world performance at low cost. This 120 billion parameter mixture-of-experts model achieves near-parity with OpenAI o4-mini on core reasoning benchmarks while running efficiently on a single 80 GB GPU. It was trained using reinforcement learning and techniques informed by OpenAI's most advanced internal models, including o3 and other frontier systems.

GPT-OSS-20B

OpenAI

GPT-OSS-20B is a medium-sized open-weight language model that delivers similar results to OpenAI o3‑mini on common benchmarks and can run on edge devices with just 16 GB of memory. This makes it ideal for on-device use cases, local inference, or rapid iteration without costly infrastructure. Despite its smaller size, it demonstrates strong performance on reasoning tasks, tool use, and competition mathematics.

Kimi K2

Moonshot AI

Moonshot's latest Mixture-of-Experts model with 32 billion activated parameters and 1 trillion total parameters. Achieves state-of-the-art performance in frontier knowledge, math, and coding among non-thinking models. Meticulously optimized for agentic tasks with sophisticated tool-use capabilities and multi-turn interactions.

Grok 4

xAI

xAI's latest generation model with enhanced mathematical reasoning capabilities, showing significant improvements in competition-level mathematics benchmarks. Features 2x faster end-to-end latency, supports 5 different voices, and achieves 10x daily user seconds compared to previous models.

Gemini 2.5 Flash-Lite

Google

Google's most cost-efficient and fastest 2.5 model yet. Higher quality than 2.0 Flash-Lite on coding, math, science, reasoning and multimodal benchmarks. Excels at high-volume, latency-sensitive tasks like translation and classification.

Claude Opus 4

Anthropic

Claude Opus 4 is Anthropic's most powerful model and the best coding model in the world, leading on SWE-bench (72.5%) and Terminal-bench (43.2%). It delivers sustained performance on long-running tasks that require focused effort and thousands of steps, with the ability to work continuously for several hours. Features include extended thinking with tool use, parallel tool execution, improved memory capabilities, and significantly reduced shortcut behaviors.

Claude Sonnet 4

Anthropic

Claude Sonnet 4 significantly improves on Sonnet 3.7's capabilities, excelling in coding with a state-of-the-art 72.7% on SWE-bench. The model balances performance and efficiency, with enhanced steerability for greater control over implementations. Features include extended thinking with tool use, parallel tool execution, improved instruction following, and significantly reduced shortcut behaviors.

Gemini 2.5 Flash

Google

Improved across key benchmarks for reasoning, multimodality, code and long context while getting even more efficient. Best for fast performance on complex tasks.

Gemini Diffusion

Google

Google's state-of-the-art, experimental text diffusion model that uses diffusion techniques to explore a new kind of language model that gives users greater control, creativity, and speed in text generation. Unlike traditional autoregressive models that generate text sequentially, diffusion models refine noise step-by-step to generate entire blocks of tokens at once.

Gemma 3n

Google

Gemma 3n is a mobile-first AI model built on a groundbreaking architecture optimized for on-device performance. It features Per-Layer Embeddings (PLE) for reduced RAM usage, allowing larger models to run on mobile devices with a memory footprint comparable to much smaller models. Supports multimodal understanding including text, images, audio, and video.

MedGemma

Google

The MedGemma collection contains Google's most capable open models for medical text and image comprehension, built on Gemma 3. Developers can use MedGemma to accelerate building healthcare-based AI applications. MedGemma comes in two variants: a 4B multimodal version and a 27B text-only version. The 4B version utilizes a SigLIP image encoder specifically pre-trained on de-identified medical data including chest X-rays, dermatology images, ophthalmology images, and histopathology slides.

Gemini 2.5 Pro

Google

Gemini 2.5 Pro is capable of reasoning through its thoughts before responding, resulting in enhanced performance and improved accuracy. Features Deep Think, an enhanced reasoning mode, and native audio outputs that capture subtle nuances of speech.

Qwen3 235B (MoE)

Alibaba

The flagship model of the Qwen 3 family, featuring a Mixture-of-Experts architecture with 235B total parameters (22B active). It introduces 'Hybrid Reasoning' capabilities, allowing it to switch between standard generation and a deep 'Thinking Mode' for complex logic. Trained on 36 trillion tokens, with support for 119 languages and dialects. Available in 6 dense models (0.6B to 32B parameters) and 2 MoE models (30B/3B active and 235B/22B active).

Qwen-3

Alibaba

Third-generation Qwen model featuring hybrid reasoning capabilities that can switch between thinking and non-thinking modes. Trained on 36 trillion tokens (double that of Qwen2.5), with support for 119 languages and dialects. Available in 6 dense models (0.6B to 32B parameters) and 2 MoE models (30B/3B active and 235B/22B active).

o3

OpenAI

OpenAI's most powerful reasoning model that pushes the frontier across coding, math, science, visual perception, and more. Sets new state-of-the-art on benchmarks including Codeforces, SWE-bench, and MMMU. Features full tool access and can agentically use and combine every tool within ChatGPT.

o4-mini

OpenAI

A smaller model optimized for fast, cost-efficient reasoning. Achieves remarkable performance for its size and cost, particularly in math, coding, and visual tasks. It is the best-performing benchmarked model on AIME 2024 and 2025, with significantly higher usage limits than o3.

GPT-4.1

OpenAI

Flagship GPT model for complex tasks. It is well suited for problem solving across domains. Features major improvements in coding, instruction following, and long context comprehension.

Llama 4 (400B)

Meta

Latest Meta AI model, at ~400B parameters and trained on 40T tokens. Intended to push open-model performance to near-GPT-4 level. Released under a new Llama 4 community license. Early indications show significant gains in reasoning and multilingual understanding.

GPT-4.1 mini

OpenAI

Smaller version of GPT-4.1 optimized for coding and instruction following while maintaining the 1M token context window. Designed for applications requiring strong coding capabilities with better efficiency.

GPT-4.1 nano

OpenAI

Most compact version of GPT-4.1 family, optimized for efficiency while maintaining strong coding and instruction following capabilities. Designed for applications requiring low latency and cost.

Qwen2.5-Omni 7B

Alibaba

An end-to-end multimodal model featuring a 'Thinker-Talker' architecture. It accepts text, image, audio, and video inputs and generates real-time streaming speech and text outputs without external ASR/TTS systems.

Gemma 3

Google

Gemma 3 is a collection of lightweight, state-of-the-art open models built from the same research and technology that powers Gemini 2.0 models. The most capable model you can run on a single GPU or TPU, delivering state-of-the-art performance for its size and outperforming Llama3-405B, DeepSeek-V3 and o3-mini in preliminary human preference evaluations on LMArena's leaderboard. Features advanced text and visual reasoning capabilities, 128K token context window, function calling, and support for over 140 languages.

GPT-4.5

OpenAI

A research preview of OpenAI's largest and most capable GPT model to date. GPT-4.5 represents a significant advancement in scaling unsupervised learning, delivering broader world knowledge, dramatically reduced hallucinations (37.1% vs 61.8% for GPT-4o), and enhanced emotional intelligence for more natural human collaboration. Excels at creative tasks, writing assistance, and complex problem-solving with improved factual accuracy and nuanced understanding of user intent.

Gemini 2.0 Flash-Lite

Google

Best for cost-efficient performance. Better quality than 1.5 Flash, at the same speed and cost. Features a 1 million token context window and multimodal input.

Gemini 2.0 Flash

Google

Next iteration of Gemini, released in three versions (Flash, Flash-Lite, Pro). Represents Google's state-of-the-art multimodal model, likely incorporating Mixture-of-Experts at unprecedented scale and targeting dominance in both text and image understanding.

Claude 3.7 Sonnet

Anthropic

Anthropic's most intelligent model to date and the first hybrid reasoning model on the market. Claude 3.7 Sonnet can produce near-instant responses or extended, step-by-step thinking that is made visible to the user. API users have fine-grained control over how long the model can think for (up to 128K tokens). Priced at $3 per million input tokens and $15 per million output tokens at launch, including thinking tokens.

Grok 3 Mini

xAI

A smaller, more efficient version of Grok 3 from xAI. Represents a new frontier in cost-efficient reasoning, particularly strong on STEM tasks that don't require as much world knowledge. Also features test-time compute and reasoning capabilities through the Grok 3 mini (Think) variant, achieving impressive performance on mathematical and coding benchmarks while being more resource-efficient than the full Grok 3 model.

Grok 3

xAI

xAI's most advanced model yet, blending superior reasoning with extensive pretraining knowledge. Trained on the Colossus supercluster with 10x the compute of previous state-of-the-art models. Features test-time compute and reasoning capabilities through reinforcement learning, allowing it to think for seconds to minutes while correcting errors and exploring alternatives. Achieved an Elo score of 1402 in the Chatbot Arena.

Gemini 2.0 Pro

Google

Google's best model yet for coding performance and complex prompts, with better understanding and reasoning of world knowledge than any previous release. Features a massive 2 million token context window and the ability to call tools like Google Search and code execution.

Qwen2.5-VL 72B

Alibaba

A state-of-the-art vision-language model capable of understanding videos over 1 hour long, performing precise object grounding, and structured output generation. It can also act as a visual agent for computer/mobile GUI operations.

DeepSeek-R1

DeepSeek

DeepSeek-R1 is a first-generation reasoning model trained via large-scale reinforcement learning. Built on DeepSeek-V3-Base, it incorporates cold-start data before RL to address challenges like endless repetition and poor readability found in DeepSeek-R1-Zero. Achieves performance comparable to OpenAI-o1 across math, code, and reasoning tasks through advanced chain-of-thought reasoning capabilities.

DeepSeek-V3

DeepSeek

A powerful Mixture-of-Experts (MoE) language model with 671B total parameters and 37B activated for each token. Features Multi-head Latent Attention (MLA) and DeepSeekMoE architectures with innovative auxiliary-loss-free load balancing and multi-token prediction training. Pre-trained on 14.8T high-quality tokens with only 2.788M H800 GPU hours.

Amazon Nova (Pro)

Amazon

Family of models (Nova Micro, Lite, Pro) launched as part of AWS's Bedrock services. Details are limited, but these models power Amazon's own generative AI offerings, similar to how Azure uses OpenAI models.

Qwen2.5-Coder 32B

Alibaba

The flagship open-weight coding model of the Qwen 2.5 series. Trained on 5.5 trillion tokens of code, it matches the coding capabilities of GPT-4o in benchmarks like HumanEval and Aider. It supports over 92 programming languages.

Mistral Large (123B)

Mistral AI

Flagship 123B-parameter model from Mistral AI, progressively upgraded through 2024 versions. Targeted to be a top-performing open model competing with Meta's Llama family on knowledge and reasoning tasks.

Pixtral 123B

Mistral AI

A 123B-parameter **multimodal** model from Mistral, combining language with vision ("Pix-"). A 12B smaller version was released under Apache 2.0. Pixtral extends Mistral's LLM capability to image understanding and description tasks.

Claude 3.5 Haiku

Anthropic

Anthropics fastest model at time of launch, delivering intelligence at blazing speeds. Claude 3.5 Haiku is optimized for quick responses while maintaining high quality.

Ministral 3B

Mistral AI

The smallest model in Mistral AI's Ministral family, part of 'les Ministraux'. A state-of-the-art edge model optimized for on-device computing with exceptional knowledge, commonsense, reasoning, and function-calling capabilities in the sub-10B category.

Ministral 8B

Mistral AI

A state-of-the-art 8B parameter model from Mistral AI's Ministral family, part of 'les Ministraux'. Features special interleaved sliding-window attention pattern for faster and memory-efficient inference, designed for edge computing and on-device applications.

Qwen2.5 32B

Alibaba

A highly efficient dense model designed to balance performance and memory requirements. It is open-sourced under Apache 2.0 license and supports long-context understanding. It outperforms many larger models (e.g., Llama-3.1-70B) in specific benchmarks while fitting on consumer-grade hardware.

Qwen2.5 72B

Alibaba

The flagship dense model of the Qwen 2.5 series, pretrained on 18 trillion tokens. It features significant enhancements in knowledge, coding, and mathematics compared to Qwen 2. It supports structured data generation (JSON) and long-context understanding.

Qwen2.5-Math 72B

Alibaba

A specialized model for mathematical reasoning, supporting Chain-of-Thought (CoT) and Tool-Integrated Reasoning (TIR). It achieves state-of-the-art performance among open-weight models on math benchmarks. It supports bilingual (English/Chinese) math reasoning.

o1-mini

OpenAI

Faster, cost-efficient reasoning model optimized for coding and agentic applications. 80% cheaper than o1-preview with strong capabilities for complex problem-solving in tasks where context is provided within the prompt.

o1-preview

OpenAI

Large reasoning model with strong capabilities for solving hard problems through extended thinking. Designed to spend more time reasoning before responding, with enhanced performance in science, coding, and math.

o1

OpenAI

The o1 model series is trained with large-scale reinforcement learning to reason using chain-of-thought. OpenAI o1 is the next model in this series (previously OpenAI o1-preview), with advanced reasoning capabilities that provide new avenues for improving safety and robustness. The model can reason about safety policies in context when responding to potentially unsafe prompts, through deliberative alignment.

Grok-2 mini

xAI

Grok-2 mini is a small but capable sibling of Grok-2 that offers a balance between speed and answer quality. It demonstrates significant improvements over Grok-1.5 while maintaining faster inference speeds compared to the full Grok-2 model.

Grok-2

xAI

xAI's Grok-2 is a significant step forward from Grok-1.5, featuring frontier capabilities in chat, coding, and reasoning. It outperformed Claude 3.5 Sonnet and GPT-4-Turbo on the LMSYS leaderboard and shows significant improvements in reasoning with retrieved content and tool use capabilities.

Llama 3.1 (405B)

Meta

Third-generation Llama model, reportedly 405B parameters with 15.6T tokens of training. Introduced a massive context window (128k tokens) and expanded multilingual capability. At release, it was the largest openly available model to date.

Claude 3.5 Sonnet

Anthropic

Claude 3.5 Sonnet raises the industry bar for intelligence, outperforming competitor models and Claude 3 Opus on a wide range of evaluations, with the speed and cost of a mid-tier model. It operates at twice the speed of Claude 3 Opus.

Qwen-2

Alibaba

Second-generation Qwen model (72B parameters) open-sourced under Alibaba's Qianwen License. Trained on 3T tokens; released alongside smaller variants and intended as foundation for Alibaba's enterprise AI services.

Codestral

Mistral AI

Mistral AI's first-ever code model. An open-weight generative AI model explicitly designed for code generation tasks. Trained on a diverse dataset of 80+ programming languages including Python, Java, C, C++, JavaScript, Bash, Swift, and Fortran. Helps developers write and interact with code through shared instruction and completion API endpoints.

GPT-4o

OpenAI

GPT-4o ('o' for 'omni') is a step towards much more natural human-computer interaction—it accepts as input any combination of text, audio, image, and video and generates any combination of text, audio, and image outputs. It can respond to audio inputs in as little as 232 milliseconds, with an average of 320 milliseconds, which is similar to human response time in a conversation. It matches GPT-4 Turbo performance on text in English and code, with significant improvement on text in non-English languages, while also being much faster and 50% cheaper in the API. GPT-4o is especially better at vision and audio understanding compared to existing models.

Mixtral 8×22B

Mistral AI

A sparse Mixture-of-Experts (SMoE) model that uses only 39B active parameters out of 141B total, offering unparalleled cost efficiency for its size. Sets a new standard for performance and efficiency within the AI community.

Phi-3

Microsoft

14B-parameter follow-up to Phi-1/2 focusing on compact "small models" with high-quality training. Released in sizes 3B, 7B, 14B. Emphasizes efficient training and inference.

Claude 3 Haiku

Anthropic

Anthropic's fastest and most compact model at the time of launch with near-instant responsiveness. Answers simple queries and requests with unmatched speed while maintaining good quality. Pricing at launch: $0.25 per million input tokens, $1.25 per million output tokens.

Claude 3 Opus

Anthropic

Anthropic's most intelligent model at release, with best-in-market performance on highly complex tasks. It can navigate open-ended prompts with remarkable fluency and human-like understanding. Pricing at launch: $15 per million input tokens, $75 per million output tokens.

Claude 3 Sonnet

Anthropic

Anthropic's balanced model that strikes an ideal balance between intelligence and speed for enterprise workloads. Delivers strong performance at a lower cost compared to its peers. Pricing at launch: $3 per million input tokens, $15 per million output tokens.

Gemini 1.5

Google

Upgrade over Gemini 1.0, featuring a **1 million+ token context window** (initially 1M, later expanded to 2M). Likely uses a Mixture-of-Experts backend and improved multimodal abilities. Released in variants (e.g. Flash).

Gemini 1.0

Google

First version of Google's Gemini, a multimodal model built by the joint Google DeepMind team. Comes in three sizes; integrated into a chatbot (Google's next-gen model after PaLM 2) with planned image and text capabilities.

Phi-2

Microsoft

2.7B-parameter model focused on efficiency and quality by training on high-quality (including synthetic) data. Achieved surprisingly strong results for its size, dubbed a "textbook-quality" model.

Mixtral 8×7B

Mistral AI

A sparse Mixture-of-Experts model (8 experts) with a total of 46.7B parameters (12.9B active per token). Outperformed dense models like Llama 2 70B and GPT-3.5 on many benchmarks at significantly lower compute cost.

Claude 2.1

Anthropic

Major update to Claude 2 featuring an industry-leading 200K token context window (roughly 150,000 words or 500+ pages), 2x reduction in hallucination rates, system prompts support, and beta tool use capabilities for API integration.

DeepSeek LLM

DeepSeek

DeepSeek LLM is an advanced language model comprising 67 billion parameters, trained from scratch on a vast dataset of 2 trillion tokens in both English and Chinese. The model demonstrates superior general capabilities, outperforming Llama2 70B Base in reasoning, coding, math, and Chinese comprehension. Available in both Base and Chat variants with 7B and 67B parameter sizes.

GPT-4 Turbo

OpenAI

Improved version of GPT-4 with significantly expanded context window, more recent knowledge cutoff, and optimized performance. Designed for production applications with better cost efficiency than the base GPT-4.

Grok 1

xAI

First model from Elon Musk's xAI, reportedly with 314B parameters. Deployed in the Grok chatbot with real-time access to X (Twitter) data. Offered an 8K context and aimed to provide up-to-date information via integration with social media.

Mistral 7B

Mistral AI

7.3B-parameter model trained from scratch with improved data and architecture enhancements. Open-source (Apache 2.0) with strong performance that rivaled larger 13B+ models on many benchmarks.

Claude 2

Anthropic

Anthropic's most capable system at release, representing a continuous evolution from previous Claude models. Trained via unsupervised learning, RLHF, and Constitutional AI with both supervised and reinforcement learning phases. Features improved helpfulness, honesty, and harmlessness (HHH) with enhanced coding abilities, longer output generation up to 4000 tokens, and better structured data formatting capabilities.

Llama 2 70B

Meta

70B-parameter successor to LLaMA, trained on 2 trillion tokens and fine-tuned for helpful dialogue. Released openly under a custom license allowing commercial use with some restrictions. Strong performance rivaling closed models like ChatGPT on many benchmarks.

GPT-3.5 Turbo 16K

OpenAI

Extended context version of GPT-3.5 Turbo with a 16K token context window. Optimized for chat applications with improved instruction following and reduced hallucinations compared to base GPT-3.5.

PaLM 2

Google

Second-generation PaLM, actually a family of models (up to ~340B dense params) with training on a diverse multilingual dataset. Powers Google's Bard and other products; improved coding, reasoning, and multilingual skills over PaLM.

Claude (v1)

Anthropic

Anthropic's first-generation AI assistant, built on research into training helpful, honest, and harmless AI systems. Claude was designed for conversational and text processing tasks with enhanced safety features, reduced harmful outputs, and improved steerability compared to earlier language models. Claude v1 introduced Constitutional AI training methods that became foundational to Anthropic's approach.

GPT-4

OpenAI

Fourth-generation GPT model, with greatly improved reasoning and multimodal capabilities (accepts text and images). Model size not publicly disclosed (rumored ~1.7 trillion parameters). Achieved human-level performance on many academic and professional benchmarks.

LLaMA 65B

Meta

Large Language Model Meta AI, released in sizes up to 65B parameters for research. Achieved strong performance by training on 1.4T tokens; weights were leaked, spurring a wave of fine-tuned variants.

AlexaTM 20B

Amazon

20B-parameter seq2seq model (teacher-student ensemble) for powering Amazon Alexa's multilingual understanding. Uses a bidirectional encoder-decoder architecture for translation and dialogue tasks.

Galactica

Meta

120B-parameter model aimed at scientific text and knowledge integration. Notorious for generating plausible-sounding but incorrect scientific content; public demo was retracted.

GPT-3.5 Turbo

OpenAI

OpenAI's public conversational AI based on GPT-3.5. Fine-tuned for dialogue via RLHF, leading to widespread adoption in late 2022. Marked the first widely accessible GPT-powered chatbot.

OPT 175B

Meta

Open Pretrained Transformer, a reproduction of GPT-3 (175B) made openly available to researchers. Notably released with training logs and a non-commercial license.

PaLM

Google

540B-parameter Transformer model trained with the Pathways system. Showed strong few-shot performance and was the largest dense model publicly disclosed as of 2022.

Chinchilla

Google

70B-parameter model trained on 4× more data than typical, to be compute-optimal. Achieved superior accuracy vs much larger models, validating data-scaling laws.

LaMDA

Google

137B-parameter dialogue-optimized language model ("Language Model for Dialogue Applications"). Known for conversational ability and for sparking discussions on AI sentience.

InstructGPT/GPT-3.5

OpenAI

Instruction-tuned version of GPT-3 (a.k.a GPT-3.5) refined via Reinforcement Learning from Human Feedback (RLHF). Powers the initial version of ChatGPT (Dec 2022). Optimized for following user instructions safely.

GLaM (Generalist Language Model)

Google

A 1.2 trillion-parameter **Mixture-of-Experts** Transformer model. Uses sparse activation (64 experts) to reduce inference cost relative to model size.

Gopher

Google

A 280B-parameter Transformer model from DeepMind, used to explore scaling laws and as a precursor to Chinchilla.

Megatron-Turing NLG 530B

Microsoft

A 530-billion-parameter Transformer language model, one of the largest dense LLMs of 2021. Collaboration between NVIDIA and Microsoft, used for research on scaling.

Codex

OpenAI

A GPT-3-descendent model fine-tuned on billions of lines of source code from publicly available sources, including code in public GitHub repositories. Powers GitHub Copilot for code autocompletion and demonstrated strong coding ability (solving ~28% of HumanEval at 12B params). Most capable in Python, but proficient in over a dozen languages including JavaScript, Go, Perl, PHP, Ruby, Swift, TypeScript, and Shell.

GPT-3

OpenAI

175B-parameter GPT series model that set new benchmarks in few-shot learning and text generation. Demonstrated impressive zero-shot capabilities across NLP tasks.

T5 (Text-To-Text Transfer Transformer)

Google

Encoder-decoder Transformer trained on a unified text-to-text format for diverse NLP tasks. Enabled prompt-based multitask learning.

XLNet

Google

Autoregressive Transformer that uses permuted language modeling as an alternative to BERT. Achieved state-of-the-art on several NLP tasks in 2019.

GPT-2

OpenAI

Generative Pre-trained Transformer 2 (GPT-2) is a large language model by OpenAI and the second in their foundational series of GPT models. It was pre-trained on a dataset of 8 million web pages (WebText) and was a direct scale-up of GPT-1 with a ten-fold increase in both parameter count and training dataset size. Initially partially withheld due to concerns about potential misuse, the full model was released in November 2019.

BERT

Google

Bidirectional Transformer pre-trained for language understanding (not generative). Influential for NLP tasks with masked-language modeling.

GPT-1

OpenAI

Generative Pre-trained Transformer 1 (GPT-1) was the first of OpenAI's large language models, demonstrating unsupervised pre-training with a 12-layer decoder-only Transformer. It introduced a two-stage training process leveraging the BookCorpus dataset and achieved improvements across various NLP tasks.