Moonshot AI

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.

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.

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.

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.

Kimi K2.5

Moonshot AI

Kimi K2.5 is a native multimodal MoE model created by continuing pretraining Kimi-K2-Base on approximately 15 trillion mixed visual and text tokens. A MoonViT vision encoder enables image and video understanding alongside coding and reasoning. Instant and thinking modes support both conversation and agent workloads. Moonshot also provides an Agent Swarm system that coordinates parallel agents; this is a harness capability around the model rather than additional model parameters. Public modified-MIT weights, an API, and Kimi chat access make the release available for research and deployment.

Kimi K2 Thinking

Moonshot AI

Kimi K2 Thinking is the reasoning specialization of the K2 trillion-parameter MoE model. It interleaves extended reasoning with tool calls and is trained to sustain long agent trajectories, with Moonshot demonstrating sequences of 200–300 tool calls. A 256K context supports research, programming, and mathematical tasks. Public weights use modified MIT terms and include native INT4 quantization. It is distinct from the original non-thinking K2-Instruct checkpoint; the reasoning and agent benchmark results belong to this model. Official API and Kimi access complement local inference.

Kimi Linear

Moonshot AI

Kimi Linear is an open-weight language-model family demonstrating a hybrid linear-attention architecture. Kimi Delta Attention provides fine-grained gated recurrent-state updates and is combined with Multi-head Latent Attention for long-context modeling. The published 48B model activates 3B parameters and supports million-token contexts, with base and instruction-tuned checkpoints trained on 5.7 trillion tokens. Moonshot releases the model and KDA kernels to support efficient long-context inference and research; efficiency measurements depend on context length and the compared implementation.

Kimi K2

Moonshot AI

Kimi K2 is Moonshot AI’s original non-thinking agentic MoE language model, with 1T total and 32B activated parameters. MuonClip-based training and large-scale agent post-training target coding, instruction following, and tool use. The July 11 release offers base and instruction checkpoints with a 128K context. The September 5 Instruct update improves coding and extends context to 256K, grouped here as a weight variant. The separately released K2-Thinking reasoning model has its own page. Public weights use Modified MIT terms; tools require an external execution harness.

Kimi-Dev

Moonshot AI

Kimi-Dev is a 72B open-weight language model specialized for software issue resolution. It builds on Qwen2.5-72B and uses reinforcement-learning-based self-play with coding and testing tasks to improve repository-level software engineering. The released checkpoint targets bug fixing and code editing under an agent harness rather than merely writing isolated snippets. MIT weights and inference examples support self-hosted experiments. End-to-end issue-resolution performance depends on the surrounding tools, tests, and execution environment.

Kimi-Audio

Moonshot AI

Kimi-Audio is a language-capable audio foundation-model family for audio understanding, text responses, and speech generation. It combines semantic and acoustic audio representations, pretraining on more than 13 million hours of speech, music, sounds, and paired text. The released 7B-base-language-model and instruction-tuned checkpoints support tasks including audio question answering, speech recognition, and spoken conversation. Because it generates and reasons with language, it is in scope beyond speech-only transcription or synthesis. Public MIT-tagged weights and inference code are available.

Kimi-VL

Moonshot AI

Kimi-VL is a compact vision-language MoE family using the native-resolution MoonViT visual encoder and a Moonlight language backbone. It supports image and video comprehension, OCR, multi-image questions, visual reasoning, and agent tasks. The language decoder activates roughly 2.8B parameters. Instruct and Thinking releases are grouped here, including the June 2025 Thinking update. Public MIT weights and inference code support local experimentation; long-thinking variants trade additional generation for more difficult reasoning.

Moonlight

Moonshot AI

Moonlight is an open-weight 16B MoE language-model family released with Moonshot research on scaling the Muon optimizer. It activates roughly 3B parameters per token and includes pretrained, instruction-tuned, and intermediate checkpoints. The research identifies optimizer adjustments for stable large-scale language-model training and provides a Muon implementation. Moonlight also serves as the language backbone for Kimi-VL. Its inclusion reflects a substantive, independently trained research model, with MIT weights available for replication and downstream studies.

Kimi K1.5

Moonshot AI

Kimi K1.5 is Moonshot’s published multimodal reasoning-model research on reinforcement learning with language models. The report emphasizes long chain-of-thought learning, context scaling, and transferring long reasoning into shorter responses. It evaluates mathematical, coding, and visual reasoning tasks and documents the training methodology. This is a historically substantive research release, not a downloadable checkpoint: no public K1.5 weights or training license are established by the report. Parameter count and the complete architecture are not disclosed sufficiently for this catalog.