Kimi K3

2026-07-16 · Publisher : Moonshot AI · Domain : Artificial Intelligence

Moonshot AI launches Kimi K3, a 2.8-trillion-parameter MoE (104 billion active) with a one-million-token context, presented as the first open model of the 3T class, with weights published on July 27, 2026.

### Context In July 2026, the closed references are Claude Fable 5 and GPT 5.6 Sol, which Moonshot AI itself acknowledges as more powerful overall. K3 succeeds K2.7 Code and K2.6 by changing architecture: it relies on Kimi Delta Attention (derived from Kimi Linear) and on Attention Residuals. ### What the model brings - Architecture change: Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), with a MoE of 896 experts of which 16 are active (Stable LatentMoE). About 2.5 times more scaling efficiency than K2, according to the vendor. - Context of 1,048,576 tokens, text, image and video input, always-on reasoning (effort low, high or max, max by default). - Training with MXFP4 quantization (weights) and MXFP8 (activations) from supervised fine-tuning onward. - Vendor demonstration cases: GPU kernel optimization, a Triton-like compiler (MiniTriton), chip design in 48 hours. - Limits stated by the vendor: sensitivity to thinking history (to be sent back in full), tendency to take too much initiative, user-experience gap with Fable 5 and GPT 5.6 Sol. ### Architecture and specifications 2.8T total parameters, 104B active, 93 layers (69 KDA and 24 gated MLA), 896 experts, 16 selected, 2 shared, hidden dimension 7168, vocabulary of 160K, MoonViT-V2 vision encoder (401 M). License: "Kimi K3 License", permissive with conditions: if the licensee or its affiliates operate a model-as-a-service (Model as a Service) business with more than 20 million dollars of revenue over 12 months, a separate agreement with Moonshot AI is required for commercial use; "Kimi K3" must be displayed beyond 100 million monthly users or 20 million dollars of monthly revenue; internal use is excluded from these constraints. Weights: huggingface.co/moonshotai/Kimi-K3. Technical report in the GitHub repository. Recommended engines: vLLM, SGLang, TokenSpeed; the vendor recommends configurations of 64 accelerators or more. ### Cost and access | Item | Price (USD per million tokens) | |---|---| | Input (cache miss) | 3.00 | | Input (cache hit) | 0.30 | | Output | 15.00 | | Cache write, 5 minutes | 3.00 | | Cache write, 1 hour | 6.00 | API identifier: kimi-k3, OpenAI- and Anthropic-compatible. Hardware: supernode of 64 accelerators or more recommended by the vendor. ### Benchmarks Scores published by the vendor (Hugging Face card, K3 at max effort), with different harnesses depending on the benchmark (Kimi Code, Claude Code, Codex). | Benchmark | Kimi K3 | Claude Fable 5 | GPT-5.6 Sol | Claude Opus 4.8 | |---|---|---|---|---| | GPQA Diamond | 93.5 | 92.6 | 94.1 | 91.0 | | DeepSWE | 67.5 | 70.0 | 73.0 | 59.0 | | Terminal-Bench 2.1 | 88.3 | 88.0 | 88.8 | 84.6 | | BrowseComp | 91.2 | 88.0 | 90.4 | 84.3 | | OSWorld-Verified | 84.8 | 85.0 | 83.0 | 83.4 | | GDPval-AA v2 (Elo) | 1686 | 1747 | 1736 | 1593 | Source: Kimi-K3 Hugging Face card; competitors' figures come from their publications or third-party rankings cited by the vendor. Fable 5 is flagged with fallback on certain tasks. ### Versus closed models The vendor itself writes that K3 remains behind Claude Fable 5 and GPT 5.6 Sol overall. The table confirms this on DeepSWE (67.5 versus 70.0 and 73.0) and GDPval-AA (1686 versus 1747 and 1736), but K3 is ahead on BrowseComp (91.2) and on par on GPQA Diamond, Terminal-Bench 2.1 and OSWorld-Verified (gaps of less than 1 point). It clearly outperforms Claude Opus 4.8 on almost every row. The price (3 and 15 USD) is published; no pricing for Fable 5 or GPT 5.6 Sol was consulted here, so the price gap is not quantified. The license requires an agreement beyond 20 million dollars of revenue for an API-type service, which limits resale as a model in service but allows internal use. ### Key takeaways Kimi K3 is the largest open model released to date and sits within a few points of the best closed models on several agent benchmarks, according to the vendor. The license and the need for a supernode matter as much as the level for an enterprise deployment.

Tags : LLM, Open Source, Moonshot AI, Kimi, MoE, Licence Kimi K3, Multimodal, Agentique