iFANN
    Last updated
    Kimi K3
    Large Language Model

    Kimi K3

    by Moonshot AI

    "The world's largest open-weight AI model, rivaling top proprietary systems."

    90%

    Rating

    4.5/5

    $3 per million input tokens / $15 per million output tokensAvailable
    Community

    Overview

    On July 16, 2026, Moonshot AI dropped a bombshell on the AI world. Kimi K3 is a 2.8-trillion-parameter open-weight model built on a sparse Mixture-of-Experts architecture with 896 experts, only 16 of which activate per forward pass. That design keeps inference costs manageable while delivering performance that, on several benchmarks, matches or beats Anthropic's Claude Fable 5 and OpenAI's GPT-5.6 Soul.

    Kimi K3 comes with a 1-million-token context window, native multimodal input for images and video, and an always-on chain-of-thought reasoning mode that cannot be disabled. Moonshot introduced two architectural innovations: Kimi Delta Attention, a hybrid linear attention mechanism, and Attention Residuals, both designed to improve information flow through very long sequences. The model is built for long-horizon agentic coding, knowledge work, and self-evolving workflows, capable of sustaining coherent multi-step technical work over extended periods.

    The hosted model went live immediately on Kimi.com, Kimi Work, Kimi Code, mobile apps, and the Kimi API. Full weights are scheduled for public release by July 27, 2026, making this the first open-source model in the 3-trillion-parameter class. API pricing is set at $3 per million input tokens and $15 per million output tokens, with cached input tokens dropping to $0.30 per million. A promotional top-up rebate running through August 12, 2026, offers up to 30% back in vouchers for API credits of $1,000 or more.

    Key Features

    • 2.8 trillion parameters with sparse MoE for efficiency
    • 1-million-token context window
    • Native multimodal input (image and video)
    • Always-on chain-of-thought reasoning
    • Kimi Delta Attention and Attention Residuals innovations
    • Open-weight release for self-hosting
    • Competitive API pricing with cached input discount

    Specifications

    Parameters2.8 trillion total, ~32 billion active
    ArchitectureSparse Mixture-of-Experts, 896 experts, 16 active per forward pass
    Context Window1 million tokens
    MultimodalNative image and video understanding
    ReasoningAlways-on chain-of-thought (effort can be set to low)
    API CompatibilityOpenAI SDK compatible

    Pros & Cons

    Pros

    • • Open-weight, can be self-hosted
    • • Top-tier coding and agentic benchmark scores
    • • Very long context window
    • • Competitive pricing vs. proprietary models

    Cons

    • • Chain-of-thought cannot be fully disabled
    • • Limited upstream capacity, frequent 429 errors on some platforms
    • • Broader reasoning and vision lag behind Claude Fable 5

    Frequently Asked Questions

    How much does Kimi K3 cost?
    Kimi K3 API pricing is $3 per million input tokens and $15 per million output tokens, with cached input tokens at $0.30 per million. There is no tiering by context length.
    Is Kimi K3 open source?
    Yes, Moonshot AI has committed to releasing Kimi K3's full weights by July 27, 2026. Until then, the model is accessible through Kimi's hosted interface, mobile apps, and API.
    How does Kimi K3 compare to Claude Fable 5?
    Kimi K3 scores 57 vs Fable 5's 60 on the Artificial Analysis Intelligence Index. K3 leads on coding and agentic benchmarks (Frontend Code Arena, SWE Marathon), while Fable 5 leads on broader reasoning and vision tasks.
    What are Kimi K3's specs?
    Kimi K3 has 2.8 trillion parameters in a sparse MoE architecture with 896 experts (16 active), a 1-million-token context window, native multimodal input, and always-on chain-of-thought reasoning.
    When was Kimi K3 released?
    Kimi K3 was announced on July 16, 2026, with the hosted model available immediately. The full model weights are scheduled for public release by July 27, 2026.

    Keywords

    Kimi K3Moonshot AIopen-weightlarge language modelMixture-of-Experts2.8 trillion parameterslong contextchain-of-thoughtcoding AIagentic AI

    Sources

    Eigent AI BlogVentureBeatBBCKimi Platform DocsMindStudio

    Product Info

    Brand
    Moonshot AI
    Category
    Large Language Model
    Released
    2026-07-16
    Price
    $3 per million input tokens / $15 per million output tokens
    Availability
    Available
    Largest open-weight model
    2.8 trillion parameters
    Context window
    1 million tokens
    Architecture
    Sparse MoE, 896 experts, 16 active
    Release date
    July 16, 2026
    Company founded
    March 2023

    Where to Buy

    Join the community

    fans discussing