Kimi K3.1 Pro Weights Go Live: a 2.8T Open Model

Moonshot AI has published the full weights of Kimi K3.1 Pro on Hugging Face, keeping a promise it made when the model reached its API on 23 October that the weights would follow “by Nov 4, 2026”. The model card calls it “our most capable 3T-class model to date”. It is a heavily post-trained version of Kimi K3, the 2.8-trillion-parameter model that went open in July as the largest open-weight release on record, and it lands while that predecessor’s numbers are still fresh: when Moonshot launched K3, it beat Claude Fable 5 on frontend code.

The architecture is a mixture-of-experts design with 2.8 trillion parameters in total and 104 billion active per token. Each layer draws on 896 experts, 16 of them selected per token plus two shared experts, across 93 layers. Attention is a hybrid: 69 layers use Moonshot’s KDA linear attention and 24 use gated DSA sparse attention, which is how the model reaches a context window of 1,048,576 tokens. A 401-million-parameter MoonViT-V2 encoder handles images, and the API quick-start guide adds video input. The weights ship in MXFP4 with MXFP8 activations, trained with quantization in the loop rather than squeezed afterwards, so the published low-precision weights are the ones the model was tuned on. Moonshot recommends vLLM, SGLang or its own TokenSpeed for serving, and Together AI is listed as a hosted provider.

The headline claim is parity with the best closed models on agentic work. On Moonshot’s own table, run at maximum reasoning effort, K3.1 Pro scores 65.7 on Terminal-Bench 4.0 against 66.4 for Claude Opus 5.5, 51.7 on SWE Marathon v1.1 against Opus’s 46.3, and 92.1 on BrowseComp, above GPT-6 Astra’s 91.5. It also posts the top score on Automation Bench-AA (73.3) and on Humanity’s Last Exam with tools (70.0). The same table is candid about the gaps. On Program Bench it manages 31.5 where Opus 5.5 reaches 65.0, on HLE-Diamond it scores 39.4 to Astra’s 65.8, and on ExploitBench it reaches 74.7 while Astra scores 100.

An independent reading points the same way. Simon Willison noted that Artificial Analysis put K3.1 Pro at 55 on its Intelligence Index, behind only Claude Opus 5.5 at 58 and ahead of Claude Fable 5.1, GPT-6 Astra and Gemini 4 Argon, all at 53, calling it the first time a soon-to-be open model had scored that high against the frontier. On Moonshot’s API it costs $2 per million input tokens and $10 per million output, the same as GPT-6.1 Sol and half the price of Opus 5.5. A Highspeed variant runs two to three times faster at twice the price. Reasoning is always on, with low, high and max effort settings and no way to switch it off, and sampling parameters such as temperature are fixed.

“Open” needs a footnote. The weights come under a custom Kimi K3.1 Pro License rather than a standard open-source licence, so commercial users should read its terms before building on it. That fits a pattern: Moonshot has sought a share of cloud revenue from providers hosting K3. Hardware is the other practical limit. Even at 4-bit, a model this size needs a multi-GPU server node to run at all, so for most developers the open weights will matter through third-party hosts and fine-tunes rather than a local install. What the release does change is the ceiling for anyone who wants to own a model outright: the strongest weights available to download now sit within a few index points of the best closed systems.