Kimi K3.1 Pro weights are out on time: 1.56 TB in MXFP4, the same license as K3, and one benchmark that doesn't match

Moonshot AI kept its date. When Kimi K3.1 Pro reached the API on October 26, the full weights were only a promise "by November 4". They are now public on Hugging Face as moonshotai/Kimi-K3.1-Pro. The repository was created privately on November 2, and the first posts linking to the download appeared on November 4, the deadline day.

What is in the repo

The model card confirms the size Moonshot described: a mixture-of-experts model with 2.8T total parameters and 104B active per token, 93 layers, 896 experts with 16 selected per token plus 2 shared, a 1,048,576-token context and a 401M-parameter MoonViT-V2 vision encoder. Compared with Kimi K3, the full-attention layers were swapped for what Moonshot calls Gated DeepSeek Sparse Attention: 69 of the 93 layers use Kimi Delta Attention and 24 use Gated DSA.

The weights ship in MXFP4, with MXFP8 activations, trained with quantization in the loop rather than quantized afterwards. Even so, the 96 safetensors shards total 1.56 TB according to the repository index. That is before any KV cache for a million-token context, so this is a model for multi-GPU servers, not for a workstation. Moonshot recommends vLLM, SGLang and TokenSpeed and links recipes for each.

The license didn't change

The license file is MIT-style with two conditions, and it is word for word the Kimi K3 license with only the model name swapped:

• A company that sells "Model as a Service", meaning API access where customers control inputs, parameters or training data, needs a separate agreement with Moonshot if its group revenue exceeds $20 million over any 12 months. Products that only embed the model inside a feature don't count.
• Commercial products with more than 100 million monthly active users or more than $20 million in monthly revenue must display "Kimi K3.1-Pro" prominently in the interface.

Internal use and access through Moonshot's own products or certified inference partners are exempt from both. So the big hosting providers are the ones who will have to talk to Moonshot; almost everyone else can download and run it.

One number doesn't match

The model card's benchmark table repeats the launch figures, including 65.7 on Terminal-Bench 4.0 against 66.4 for Claude Opus 5.5 and 58.2 for GPT-6 Astra, and the 81.0% pass@3 on the internal xdev-bench-hard security test that our first post covered. All of these are company-reported.

One row can be checked against its source already. The card lists Kimi K3.1 Pro at 69.4 on Mercor's APEX-Agents and says the scores are "cited from the official leaderboard". Mercor's own announcement on November 5 puts it at 65.5% pass@1, second among open-weight models behind DeepSeek-V4.1-Flash-1017 at 69.8%.

The DeepSeek figure matches in both places; the Kimi one is four points lower in Mercor's run. The card may have been written against an earlier leaderboard entry or a different setting, but as it stands the independent number is the lower one. Moonshot's own table also shows where the gap to closed models remains, which matches the 12-point open-versus-closed gap Artificial Analysis reported last week: on HLE-Full without tools, Kimi K3.1 Pro scores 51.1 against 59.8 for Claude Opus 5.5.

What's still open is whether the downloadable model behaves like the hosted one. Moonshot's card reports the same 81.0% security score it gave for the filtered production model, but independent runs on the public weights haven't been published yet.