{"slug": "kimi-k3-moonshot-ai-s-2-8t-parameter-open-weight-frontier-model", "title": "Kimi K3: Moonshot AI's 2.8T-Parameter Open-Weight Frontier Model", "summary": "Moonshot AI released Kimi K3, an open-weight 2.8-trillion-parameter Mixture-of-Experts model with 104B activated parameters, built on the new Kimi Delta Attention architecture and featuring a 1,048,576-token context window and native multimodal vision. The company claims it is the world's first open 3T-class model, with self-reported benchmarks showing it leads on agentic tool-use and vision tasks but trails Claude Fable 5 and GPT-5.6 Sol on some reasoning evals. Weights are available on Hugging Face under the custom Kimi K3 License.", "body_md": "# Kimi K3: Moonshot AI's 2.8T-Parameter Open-Weight Frontier Model\n\nKimi K3 is Moonshot AI's open-weight successor to Kimi K2 — a 2.8T-parameter, 104B-activated MoE model built on a new Kimi Delta Attention (KDA) architecture, with native multimodal vision, a 1,048,576-token context window, and published benchmarks against Claude and GPT-class frontier models.\n\n- ⭐ 8176\n- Python\n- Kimi K3 License\n- Updated 2026-08-08\n\n[1M Context Window LLM 2026: Real Tests](https://dibi8.com/resources/llm-frameworks/1m-context-window-llm-2026-real-test/) •\n[Open Interpreter: A Codex Fork That Emulates Claude Code and Kimi’s Harness](https://dibi8.com/resources/llm-frameworks/openinterpreter-low-cost-model-coding-agent-2026/)\n\n*Kimi K3 — from github.com/MoonshotAI/Kimi-K3*\n\n## What Is Kimi K3? [#](#what-is-kimi-k3)\n\n**Kimi K3** is Moonshot AI’s newest open-weight model, and per the README, their most capable to date: a **2.8-trillion-parameter** Mixture-of-Experts model with **104B activated parameters** per token, built on a new **Kimi Delta Attention (KDA)** and **Attention Residuals (AttnRes)** architecture. Moonshot AI describes it as “the world’s first open 3T-class model” — native multimodal (text, image, video-frame understanding), a **1,048,576-token context window**, and released with full weights under a custom **Kimi K3 License**.\n\n🔗 **GitHub**: [https://github.com/MoonshotAI/Kimi-K3](https://github.com/MoonshotAI/Kimi-K3)\n🤗 **Weights**: [huggingface.co/moonshotai/Kimi-K3](https://huggingface.co/moonshotai/Kimi-K3)\n📄 **Tech report**: linked from the repo as `k3_tech_report.pdf`\n\nAt **8,100+ GitHub stars**, created July 27, 2026 and pushed as recently as August 6, 2026.\n\n## Architecture: What Changed From K2 [#](#architecture-what-changed-from-k2)\n\n| Spec | Value |\n|---|---|\n| Total parameters | 2.8T |\n| Activated parameters | 104B |\n| Layers | 93 (1 dense + 92 MoE) |\n| Attention composition | 69 KDA + 24 Gated MLA |\n| Attention heads | 96 (hidden dim 7168) |\n| Experts | 896 total, 16 selected/token, 2 shared |\n| Vocabulary | 160K tokens |\n| Context length | 1,048,576 tokens |\n| Vision encoder | MoonViT-V2 (401M params) |\n| Quantization | Native MXFP4 weights / MXFP8 activations (quantization-aware trained) |\n\nPer Moonshot AI, the **Stable LatentMoE** framework activating 16 of 896 experts yields “an approximate 2.5× improvement in overall scaling efficiency” over K2. The quantization detail is worth flagging separately: MXFP4/MXFP8 is trained in **from the SFT stage onward**, not bolted on as a post-hoc compression step — the stated goal being broad hardware compatibility without a separate quantization pass degrading quality.\n\n## Benchmark Highlights (Self-Reported, Max Effort) [#](#benchmark-highlights-self-reported-max-effort)\n\nMoonshot AI’s README publishes a large comparison table against Claude Fable 5, Claude Opus 4.8, GPT-5.6 Sol, GPT-5.5, and GLM-5.2. A representative slice — K3 doesn’t sweep every category, and the pattern shifts by benchmark type:\n\n| Benchmark | Kimi K3 | Best of the rest |\n|---|---|---|\n| BrowseComp (agentic web) | 91.2 | GPT-5.6 Sol 90.4 |\n| MCPMark-Verified (MCP tool use) | 94.5 | GPT-5.6 Sol / GPT-5.5 tied 92.9 |\n| Terminal-Bench 2.1 | 88.3 | GPT-5.6 Sol 88.8 |\n| GPQA Diamond (reasoning) | 93.5 | GPT-5.6 Sol 94.1 |\n| HLE-Full | 43.5 | Claude Fable 5 53.3 |\n| CritPt (physics reasoning) | 23.4 | GPT-5.6 Sol 32.3 |\n| Video-MME (w/ subtitles) | 90.0 | GPT-5.6 Sol 89.5 |\n| OmniDocBench (document vision) | 91.1 | Claude Fable 5 89.8 |\n| Harvey Lab-AA (legal) | 94.6 | Claude Fable 5 93.6 |\n\n**Read this carefully**: these are Moonshot AI’s own numbers from their tech report, not third-party reproductions. K3 leads on agentic tool-use and several vision/document benchmarks, but trails Claude Fable 5 and GPT-5.6 Sol on some of the hardest pure-reasoning evals (HLE-Full, CritPt). Treat it as “strong, benchmark-dependent,” not “best at everything.”\n\n## Deployment and Model Usage [#](#deployment-and-model-usage)\n\nRecommended inference engines, per the README:\n\n— published recipes at recipes.vllm.ai[vLLM](https://github.com/vllm-project/vllm)— cookbook at docs.sglang.io[SGLang](https://github.com/sgl-project/sglang)**TokenSpeed**— recipes at lightseek.org\n\nA hosted, OpenAI/Anthropic-compatible API is available at platform.kimi.ai (model name `kimi-k3`\n\n).\n\n**Thinking is always on.** Reasoning effort is set via a `reasoning_effort`\n\nfield (`\"low\"`\n\n/ `\"high\"`\n\n/ `\"max\"`\n\n, default `\"max\"`\n\n), returned as a separate `reasoning_content`\n\nfield. The one gotcha worth flagging for anyone integrating this: K3 was trained in **preserved thinking history mode**, meaning multi-turn calls must pass the complete prior assistant message back — `reasoning_content`\n\nand `tool_calls`\n\nincluded, not just the final `content`\n\nstring — or the model loses the thread on follow-up turns.\n\nFor an agent harness, Moonshot AI points to their own **Kimi Code CLI** — run it in a terminal and switch to K3 with the `/model`\n\ncommand.\n\n## Licensing: What Triggers a Paid Agreement [#](#licensing-what-triggers-a-paid-agreement)\n\nThe **Kimi K3 License** (custom, same shape as K2’s) is permissive by default — free to use, modify, fine-tune, and redistribute — with two revenue-gated conditions:\n\n**“Model as a Service” threshold**: if you (plus affiliates) give third parties API-level control over K3’s inputs/parameters/fine-tuning and your combined revenue exceeds**$20M USD over any 12 consecutive months**, you need a separate commercial agreement with Moonshot AI.** Attribution at scale**: if K3 powers a product with** 100M+ monthly active users**or**$20M+/month revenue**, “Kimi K3” must be prominently displayed in that product’s UI.\n\nBoth conditions are waived for internal use and for end-user products that merely embed K3’s capabilities without exposing model-level control to third parties.\n\n## License [#](#license)\n\n**Kimi K3 License** (custom, permissive with revenue-based commercial terms) — see [LICENSE](https://github.com/MoonshotAI/Kimi-K3/blob/main/LICENSE).", "url": "https://wpnews.pro/news/kimi-k3-moonshot-ai-s-2-8t-parameter-open-weight-frontier-model", "canonical_source": "https://dibi8.com/resources/llm-frameworks/kimi-k3-open-frontier-model-2026/", "published_at": "2026-08-07 15:00:00+00:00", "updated_at": "2026-08-09 13:00:23.501758+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "generative-ai", "ai-research", "ai-products"], "entities": ["Moonshot AI", "Kimi K3", "Kimi K2", "Kimi Delta Attention", "Claude Fable 5", "GPT-5.6 Sol", "Hugging Face", "GLM-5.2"], "alternates": {"html": "https://wpnews.pro/news/kimi-k3-moonshot-ai-s-2-8t-parameter-open-weight-frontier-model", "markdown": "https://wpnews.pro/news/kimi-k3-moonshot-ai-s-2-8t-parameter-open-weight-frontier-model.md", "text": "https://wpnews.pro/news/kimi-k3-moonshot-ai-s-2-8t-parameter-open-weight-frontier-model.txt", "jsonld": "https://wpnews.pro/news/kimi-k3-moonshot-ai-s-2-8t-parameter-open-weight-frontier-model.jsonld"}}