Kimi K3.1 Pro open weights: 2.8T MoE, 1M context, licence terms and Moonshot's own benchmarks
Moonshot has published the weights of its 2.8-trillion-parameter model on Hugging Face under a custom licence; every benchmark so far is the company's own.
· 6 min read
Written by AI from the sources linked below. How we use AI
On this page
Moonshot AI has put the weights of Kimi K3.1 Pro on Hugging Face, a 2.8-trillion-parameter mixture-of-experts model with a 1-million-token context window. The model has been on Moonshot's API since late October; the open release adds a custom licence that is close to MIT but with two commercial conditions, and a stated effort to hold back cyber capability. The benchmark numbers below are all Moonshot's own.
Key points
- The Hugging Face repository
moonshotai/Kimi-K3.1-Proholds 96 safetensors shards (about 1.56 TB in total), with a model card that calls it "open-weight" and "natively multimodal" (text and image input). - It has 2.8T total parameters, 104B active per token, 896 experts with 16 selected per token, and a context length of 1,048,576 tokens, with weights stored in MXFP4.
- The licence is custom (
license: other): broad use, modification and redistribution rights, but a separate agreement is needed for model-as-a-service businesses above US$20 million in revenue, and a visible credit for very large products. - On Moonshot's own tables it nearly matches Claude Opus 5.5 on Terminal-Bench 4.0 (65.7 against 66.4) and trails it by a wide margin on Program Bench (31.5 against 65.0).
- Moonshot says it removed cyber-related data before release; an unmitigated checkpoint solved 90.5% of an exploit benchmark, the released one 81%.
What happened
Moonshot introduced Kimi K3.1 Pro through its apps and API in the last week of October 2026. Coverage disagrees slightly on the day: Let's Data Science dates the API release to 23 October, and VKTR's article was published on 25 October. The company said at launch that the weights would follow by 4 November, which is today. By the time of the Hacker News thread on 4 November, with 79 points when we looked, the files were available.
We checked the repository directly. Its file list contains the LICENSE, the model card, configuration and custom Python code for the model and its vision processor (kimi_k3_processor.py, kimi_k3_vision_processing.py), and 96 weight shards. Hugging Face's API reports the repository's total storage at 1,561,355,732,732 bytes. We did not download or run the model, so everything about its behaviour comes from Moonshot's model card and from the press coverage cited below.
The model card describes K3.1 Pro as "our most capable 3T-class model to date", post-trained from Kimi K3 with large-scale reinforcement learning "across hundreds of downstream domains", and the first model in a K3.1 series. It lists improvements in coding, 3D creation, financial research and factual accuracy. Moonshot also sells a faster variant, K3.1 Pro Highspeed, which VKTR reports as 2-3 times faster. That variant is an API product; we found no weights for it.
Background
Moonshot introduced the first Kimi K3 in July, according to VKTR, and K3.1 Pro is a post-trained version of it. Search coverage we found says K3's weights and technical report went up together on Hugging Face and GitHub under a bespoke "Kimi K3 License"; we did not verify that against Moonshot's own pages.
The licence matters because for companies the question is less whether they can download an open-weight model than whether they can build a paid product on it. Terms like Moonshot's decide that.
How it works
The model card's architecture table gives the main design choices:
- Mixture of experts. 896 experts per layer, 16 chosen for each token, plus 2 shared experts. Only 104B of the 2.8T parameters are active for any one token, which is why a model this large can run at all.
- Attention. 93 layers: 69 use KDA attention, 24 use "Gated DSA". Moonshot says K3.1 replaces K3's full-attention Gated MLA layers with Gated DeepSeek Sparse Attention for more efficient long-context work.
- Context. 1,048,576 tokens, with a 160K-token vocabulary.
- Vision. A 401M-parameter encoder called MoonViT-V2 handles images; the model takes text and images in and produces text.
- Precision. Weights in MXFP4 and activations in MXFP8, trained with quantization-aware training, which is why the checkpoint is roughly 1.56 TB rather than several times that.
Running it. The model card links recipes for vLLM and SGLang. Our own arithmetic, not Moonshot's: the weights alone take about 1.56 TB before any context cache, so this is a job for a large multi-GPU node or several nodes, not a workstation. Most people will use it through an API or a certified inference partner.
The numbers
Everything here is from Moonshot's model card, run at "reasoning effort max", and The comparison models are labelled "max, w/ fallback" in the table. Several benchmarks are internal. Treat them as claims until independent evaluators publish their own runs.
The picture is mixed. K3.1 Pro leads its table on SWE Marathon v1.1 (51.7, against 46.3 for Opus 5.5 and 42.5 for GPT-6 Astra), on BrowseComp (92.1 against 91.5 and 90.8) and on HLE-Full with tools (70.0 against 67.7 for Opus 5.5). It is far behind on Terminal-Bench-Science 0.1 (38.1, against 63.3 for Opus 5.5 and 68.1 for GPT-6 Astra) and on Program Bench. Moonshot's hallucination figure, per VKTR, is 8.3% on its Kimi Hallu Bench Hard with tools, down from 23.4% for K3, while GPT-6 Astra scores 5.5%. On price, Let's Data Science, citing Simon Willison, reports US$2 per million input tokens and US$10 per million output tokens on the API; we could not confirm that on Moonshot's own page.
The licence in plain words
The text of the LICENSE file is short. It grants the right "to use, copy, modify, merge, publish, distribute, sublicense, and/or sell" the software, including weights, with these conditions:
- Keep the copyright notice and follow applicable laws.
- If you or an affiliate run a "Model as a Service" business, meaning you give third parties inference or fine-tuning access with meaningful control over inputs, parameters or training data, and your group's revenue exceeds US$20 million over any 12 consecutive months, you need a separate agreement with Moonshot before commercial use.
- If a commercial product built on the model has more than 100 million monthly active users or more than US$20 million in monthly revenue, "Kimi K3.1-Pro" must be shown prominently in its interface.
- Neither condition applies to internal use that does not expose the model or its outputs to third parties, or to use through Moonshot's official products or certified inference partners.
For a startup, a researcher or a company using the model inside its own tools, that is effectively permissive. For a cloud provider reselling hosted inference above the revenue threshold, it is a restricted licence, and Hugging Face tags it simply as "other".
What it means
For developers: a 104B-active model with a million-token context and a permissive licence is a new option for self-hosted coding agents, if you have the hardware. Because the weights are large, most early use will go through hosted partners.
For competitors: the mixed benchmark profile suggests Moonshot is tuning for agentic coding and search tasks, not for everything. The gaps on Program Bench and the science terminal tasks show where closed models still lead on Moonshot's own numbers.
For safety policy: the cyber note is the most interesting part of the card. Per VKTR, Moonshot removed cyber-related data from the general mid-training corpus and cyber tasks from its reinforcement-learning environments before the open release. On xdev-bench-hard, a set of exploit tasks, an unmitigated checkpoint solved 90.5%, the released model 81%, and K3 33.3%. In other words the mitigation cut the score but the released model is still far above its predecessor. Moonshot says K3.1 Pro stays behind frontier proprietary models that have been public for at least three months, and plans to raise the cyber capability available in open weights gradually.
What we don't know yet
- Independent benchmark results. All the figures above come from Moonshot.
- The real cost of self-hosting, and which inference partners are certified under the licence.
- How the cyber mitigation holds up against fine-tuning: the licence allows modification, and the card does not say whether the safeguards survive it.
- Whether the Highspeed variant will ever be open.
- The API price on Moonshot's own page.
What to watch next
Independent evaluations on public leaderboards, the first vLLM and SGLang throughput reports on real hardware, Moonshot's list of certified inference partners, and whether other open-weight labs copy the staged approach to cyber capability.
Sources
Written by
The official account of the people who run audiotale.net: announcements, conversation starters and the newsroom's drafts. Posts written with AI say so. Every other account here is a real person.
Read next
Reflection AI releases Beam, a 501B-parameter open-weight model, under Apache 2.0
The US lab says its mixture-of-experts coding and agent model is now downloadable, but every benchmark so far is its own.
audiotale.net team5 min read
How to run an open-weight LLM locally: GGUF, quantization, Ollama and llama.cpp
A step-by-step guide to running an open-weight model on your own computer: what GGUF is, how quantization sets memory needs, and the Ollama and llama.cpp commands.
audiotale.net team6 min read
Hugging Face Model Integration Contracts: tiny CPU fixtures caught the UMT5 attention bug that lived in Transformers main for two months
Hugging Face proposes pinned, shrunk model fixtures that run on cheap CPUs for every pull request, after a UMT5 regression went unnoticed from 30 July to 28 September 2026.
audiotale.net team7 min read