MoonshotAI/Kimi-K3

Open Frontier Intelligence

What it solves

Kimi K3 provides frontier-level intelligence for complex, long-horizon tasks that typically require significant human oversight. It specifically targets high-end engineering, deep research, and multimodal knowledge work, such as GPU kernel optimization, chip design, and video editing.

How it works

Kimi K3 is a 2.8-trillion parameter Mixture-of-Experts (MoE) model. It utilizes a novel architecture featuring Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), alongside a Stable LatentMoE framework that activates 16 out of 896 experts per token. It is natively multimodal, integrating text, images, and video processing within a single model, and supports a massive 1-million-token context window.

Who it’s for

It is designed for researchers, developers, and engineers who need an open-weight model capable of navigating massive code repositories, orchestrating terminal tools, and performing deep multimodal research.

Highlights

  • Massive Scale: The first open 3T-class model with 2.8 trillion total parameters.
  • Native Multimodality: Unified understanding of text, images, and video.
  • Long-Horizon Capabilities: Capable of sustaining long engineering sessions and complex software development (e.g., CAD, compiler development).
  • Extreme Context: Supports a context window of up to 1 million tokens.
  • Efficient Scaling: Uses a Stable LatentMoE framework to improve scaling efficiency by approximately 2.5× over its predecessor.

Related

  • Dispatch
    Kimi K3 Architecture Overview

    Kimi K3 is a 2.8T parameter open-weight model that optimizes inference efficiency through LatentMoE, Multi-Head Latent Attention, and a total removal of positional embeddings (NoPE).

  • Dispatch
    Moonshot AI Kimi-K3 Release

    Moonshot AI has released Kimi-K3, the first open-weights 3T-class model designed for frontier intelligence in coding, reasoning, and long-horizon knowledge work.

  • Project
    MoonshotAI/Kimi-K2.5

    Kimi K2.5 is an open-source, native multimodal agentic model with a 1T parameter MoE architecture that integrates vision and language understanding for complex task execution.

  • Dispatch
    Kimi K3 Release: 2.8T Open-Weight Model with 1M Context, Agentic Capabilities, and Frontier Benchmark Scores

    Kimi K3, a 2.8‑trillion‑parameter open‑weight model with 1 M token context and native vision, launched with pricing comparable to Anthropic’s Sonnet and benchmark results that place it just behind Claude Fable 5 and GPT‑5.6 Sol on agentic tasks.

  • Dispatch
    vLLM Kimi K3 Support

    vLLM has released day-0 support for Kimi K3, a 2.8-trillion-parameter multimodal MoE model, featuring optimizations for Kimi Delta Attention and DSpark speculative decoding to achieve up to 370 tok/s.