A 2.8-trillion-parameter model you can download
Moonshot AI has published the weights for Kimi K3.1 Pro, its largest model, on Hugging Face. The repository holds the model in 96 safetensors shards and is not gated: there is no application and no approval step between a download command and the file. Moonshot had said the weights would follow the model’s launch in late October, and they are now public.
The company calls it its “most capable 3T-class model to date”. It is a mixture-of-experts design with 2.8 trillion total parameters, of which 104 billion are activated for any given token, and a context window of 1,048,576 tokens.
What is in the file
The architecture runs to 93 layers, one of them dense, with 896 experts of which 16 are selected per token alongside two shared experts. Attention is split 69 Kimi Delta Attention layers to 24 Gated DeepSeek Sparse Attention layers — a change from Kimi K3, which used full-attention Gated MLA layers, and one Moonshot says makes long-context modelling more efficient. The attention hidden dimension is 7,168 across 96 heads, and the vocabulary runs to 160,000 tokens.
The model takes text and images; a MoonViT-V2 vision encoder of 401 million parameters sits alongside the language stack. Weights ship quantised to MXFP4 with MXFP8 activations, trained that way rather than compressed afterwards, and Moonshot points serving users at vLLM and SGLang. K3.1 Pro was post-trained from Kimi K3 with what the company describes as large-scale joint reinforcement learning across hundreds of downstream domains.
The scores, and where they come from
Moonshot publishes a comparison table against Claude Fable 5.1, GPT‑6 Astra, Claude Opus 5.5, DeepSeek-V4.1-Flash-1017 and GLM-5.3. The footnotes matter: Moonshot ran K3.1 Pro itself, in most rows through the Claude Code harness, while the competing figures are the maximum values those vendors reported in their own blog posts, system cards and leaderboard entries. It is a table of vendor claims, assembled by one of the vendors.
On that basis K3.1 Pro scores 65.7 on Terminal-Bench 4.0 against 66.4 for Claude Opus 5.5 — second by 0.7 of a point, and ahead of GPT‑6 Astra at 58.2 and Claude Fable 5.1 at 57.9. It leads on SWE Marathon v1.1 with 51.7 against Opus 5.5’s 46.3, and on BrowseComp with 92.1 against Astra’s 91.5.
It does not lead everywhere. DeepSeek-V4.1-Flash-1017 takes DeepSWE v1.1 with 77.0 to K3.1 Pro’s 75.5, and GPT‑6 Astra takes GPQA-Diamond at 96.1 against 94.2. On Humanity’s Last Exam the gap is wider: K3.1 Pro scores 51.1 where Opus 5.5 reaches 59.8 and Claude Fable 5.1 reaches 57.6, which puts the open model last among the frontier systems in that row. On OSWorld V2.1 it trails at 79.6, behind Opus 5.5 at 81.8 and Fable 5.1 at 80.7.
The licence is the catch
The weights are downloadable, but open-weight is not the same as open source. K3.1 Pro is released under the Kimi K3.1 Pro License, a document the company wrote for this model rather than Apache 2.0 or MIT, and Hugging Face records the licence field simply as “other”. Anyone planning to build on the model has to read that text rather than rely on terms they already know.
What to watch
The useful number will be the first independent one. Every figure in Moonshot’s table is a vendor’s own, Moonshot’s included, and Terminal-Bench keeps a public leaderboard that the company already cites for its rivals. A K3.1 Pro entry there, run by someone else, would settle whether the 0.7-point gap to Claude Opus 5.5 survives a harness Moonshot did not choose.
