Moonshot AI published the full weights of Kimi K3.1 Pro on Hugging Face on November 4, 2026, meeting the deadline it set when it launched the model on October 23. The release makes one of the largest open-weight models ever published freely downloadable, and it arrives at a moment when the gap between open models and the best closed models is narrowing to a few percentage points on the hardest tests.
Kimi K3.1 Pro is a mixture-of-experts model with 2.8 trillion total parameters, of which about 104 billion are active for each token, selected from 896 experts. It accepts text and image input, keeps the one-million-token context window introduced with Kimi K3, and is aimed squarely at agentic work — long, multi-step tasks carried out by AI systems with tools. The weights are stored at 4-bit precision after quantization-aware training, and the download is about 1.56 terabytes. A full-precision copy would be roughly 5.6 terabytes. Moonshot published deployment recipes for the vLLM and SGLang inference engines alongside the weights.
What the Model Can Do
Moonshot launched Kimi K3.1 Pro as the first model in its new K3.1 series and its most capable open-weight model to date. It is a post-trained version of Kimi K3, built with what the company calls substantially larger post-training compute, including large-scale joint reinforcement learning across hundreds of tasks. It is live in Kimi.ai, Kimi Work, Kimi Code, and the Kimi API, alongside a faster Kimi K3.1 Pro Highspeed variant that Moonshot says runs two to three times faster.
On Terminal-Bench 4.0, a test that measures coding agents on real terminal tasks, Moonshot reports Kimi K3.1 Pro scoring 65.7 percent, against 66.4 percent for Anthropic's Claude Opus 5.5 and 58.2 percent for OpenAI's GPT-6 Astra. Those are Moonshot's own evaluations, and independent reproductions have not yet been reported, so the numbers should be read as the company's claim rather than a settled result. Moonshot also reports 94.2 on GPQA-Diamond, 92.1 on BrowseComp, and 79.6 on OSWorld V2.1. The hosted model is priced at $2 per million input tokens and $10 per million output tokens.
The Architecture Inside
Two technical details stand out. First, in 24 of the model's 93 layers, full attention is replaced by Gated DSA, a sparse attention mechanism derived from DeepSeek's sparse attention work, which speeds up processing for very long contexts. Second, the model includes a 401-million-parameter vision encoder, MoonViT-V2, giving it native multimodal input. It also uses a custom activation function called SiTU-GLU. These are the kinds of engineering choices that matter to the small number of organizations that will actually run a model of this size — and Moonshot's own guidance makes clear that running it yourself is not trivial. The simplest vLLM setup needs eight data-center GPUs at the B300 or GB300 level, so most users will keep using the hosted API.
The Licence and the Cyber Restriction
The licence is the part founders and lawyers should read most carefully. It is a custom licence, in the style of the Kimi K3 licence, that is free to use, modify, and redistribute — but with a condition: a company running a model-as-a-service business with more than $20 million in revenue over 12 months must sign a separate agreement with Moonshot before using it commercially. That revenue gate means the model is genuinely open for startups and researchers, while large commercial hosts need a deal. Anyone planning a hosted product on Kimi K3.1 Pro should confirm the licence terms before committing.
There is also a first for Moonshot in the training itself. The company says it deliberately removed cyber-related data from mid-training and reinforcement learning to weaken the model's cyber capabilities, and committed not to release open weights that would extend the absolute frontier of cyber capability. That is a safety choice with a commercial cost: a model deliberately made weaker in one domain. It also reflects the growing pressure on open-weight releases, where the most capable models raise the hardest questions about misuse.
Why It Matters for Founders and Investors
The release is a landmark for open-weight AI. A model with 2.8 trillion parameters, a million-token context, and near-frontier coding performance is now downloadable by anyone, with deployment recipes included. Along with Tencent's Hy4 the same week, it shows that the frontier is no longer a closed club: the best open models are within a point or two of the best closed ones on key benchmarks.
For founders, the practical effect is cheaper, more capable building blocks. A startup can now build an agentic product on a model it can inspect, fine-tune, and self-host, rather than renting intelligence by the token from a single vendor. But the licence's revenue gate and the sheer hardware cost of self-hosting mean the real winners may be the infrastructure companies — the GPU clouds and inference providers — that make running these giants practical. The model is open; running it at scale is still a business of its own.
Angeled.com connects founders with investors.
Sources: Hugging Face model card, moonshotai/Kimi-K3.1-Pro, November 4, 2026; Moonshot AI tech blog, October 23, 2026; ExplainX, November 4, 2026; Daily.dev, November 5, 2026; MLLM.io, November 4, 2026.
Read Next
OpenAI Releases Codex Security as Open-Source CLI and SDK
Databricks Unveils Always-On Genie, an AI Agent That Stops Waiting to Be Asked
DeepMind Launches AlphaGenome 2, a DNA Model That Reads Genomes Cell by Cell
ElevenLabs Unveils 13 New Products and Upgrades at Its New York Summit
Google Launches Gemini Decision 1, a Model That Answers Instead of Writing
Runway Open-Sources Praxis-1, Its Robot Model That Learned by Watching Video



