Moonshot AI put the weights of Kimi K3.1 Pro on Hugging Face on November 4, 2026, twelve days after the model went live in its own API. The file set describes a mixture-of-experts network with 2.8 trillion total parameters and 104 billion active per token, a context window of 1,048,576 tokens, and text and image input. Anyone can download it. Whether anyone can build a business on it depends on a license that most coverage of the release skipped.
Video: Kimi K3.1 Pro tested against Claude Opus 5.5, from the WorldofAI channel on YouTube. The channel runs its own tests; the figures in this article come from the sources listed below.
What Moonshot published
The release is a post-trained version of Kimi K3, with, in Moonshot’s words, substantially larger post-training compute. The model card lists 93 layers, one of them dense, with 896 experts of which 16 are selected for each token plus 2 shared experts. Attention is split between 69 KDA layers and 24 gated DSA layers. Weights ship in MXFP4 with MXFP8 activations, produced with quantization-aware training, and the vision encoder is a 401-million-parameter module called MoonViT-V2.
What the card leaves blank matters just as much. It names vLLM, SGLang and TokenSpeed for serving but gives no GPU or memory requirement, and it does not say whether the faster Highspeed variant sold through the API will ever get weights. TechNode reported that Huawei has announced inference support through vLLM Ascend and SGLang on its Atlas A3 and Ascend 950 systems, so at least one non-Nvidia path exists on day one.
Where the model came from
Kimi K3.1 Pro continues a size progression at Moonshot. Build Fast with AI’s July 2026 review of K3 lists Kimi K2.7 Code as a 1-trillion-parameter, 256,000-token coding specialist with open weights, and K3 as a 2.8-trillion-parameter flagship with a 1-million-token window and text, image and video input. K3.1 Pro keeps K3’s shape and changes the post-training, aimed at coding, 3D creation, financial research and long-running agent tasks, according to TechNode.
Before the weights appeared, the model was reachable through Kimi.ai, Kimi Work, Kimi Code and the Kimi API. Analytics Vidhya reported a Highspeed variant, sold under the API names kimi-k3.1-pro and kimi-k3.1-pro-highspeed, that runs two to three times faster. The model card does not say whether weights for that faster variant will follow.
Open weights are not open source
A model counts as open weight once its trained parameters can be downloaded, whatever the license says. BenchLM draws that line plainly, and the Open Source Initiative draws a stricter one. Version 1.0 of the Open Source AI Definition, published on October 28, 2024, lists four freedoms: to use, study, modify and share a system for any purpose. Exercising them requires the preferred form for making modifications, which the definition spells out as enough information about the training data to rebuild a substantially equivalent system, the complete code used to train and run it, and the parameters themselves.
Kimi K3.1 Pro supplies the third item. The model card describes neither the training data nor the training code, which is the normal pattern for large releases. BenchLM’s catalog, as of October 29, held 528 open-weight entries among 1,009 models; a license was recorded for only 152 of them, and 107 of those 152 used Apache 2.0 or MIT. The remaining models sit under custom terms such as the Llama Community License, the Qwen Community License and now the Kimi licenses.
The two clauses that decide who can use it
The Kimi K3.1-Pro License grants the right to use, copy, distribute, sell, deploy, modify and fine-tune the software, and the conditions apply to derivative works as well. Two sections carry the restrictions.
- Model-as-a-Service (Section 2). The clause covers giving a third party access to inference or fine-tuning, an API for example, in a way that lets them meaningfully control inputs, parameters or training data. A licensee or affiliate running such a business with aggregate revenue above US$20 million over any 12 consecutive months needs a separate agreement with Moonshot before using the software or its derivatives commercially. Products that embed the model inside specific features, and services that merely relay requests to models hosted elsewhere, are excluded.
- Display (Section 3). When the software or a derivative is used in a commercial product with more than 100 million monthly active users, or more than US$20 million in monthly revenue, the name “Kimi K3.1-Pro” has to be prominently displayed in the product’s user interface.
Section 4 removes both obligations for internal use that does not expose the model or its outputs to third parties, and for access through Moonshot’s own products or certified inference partners. In practice a startup fine-tuning the model for its own staff owes nothing, a mid-sized company embedding it in a product owes nothing until the revenue and user thresholds are hit, and a hosting company reselling API access is the one the license is written for. The thresholds use different clocks, a 12-month total in one place and monthly figures in the other, which is easy to misread.
The license is not legal advice and the text is Moonshot’s; a company near either threshold would want counsel to read it, and Moonshot lists license@moonshot.ai for questions.
How the performance claims hold up
The headline numbers on the model card are Moonshot’s own. Against Claude Opus 5.5 they are mixed: Terminal-Bench 4.0 comes in at 65.7 against 66.4 for Opus, SWE Marathon v1.1 at 51.7 against 46.3 in K3.1 Pro’s favor, and then Program Bench at 31.5 against 65.0 and Terminal-Bench-Science at 38.1 against 63.3, both well behind. Several of the benchmarks are internal and have no outside replication.
An independent measurement exists for the launch version. Artificial Analysis scored K3.1 Pro at 55 on its Intelligence Index, fourth overall, up 11 points from Kimi K3’s 44 and behind Claude Opus 5.5 at 58, Claude Sonnet 5.5 at 56 and Muse Wonder Full at 56. It put the gap between the leading open-weight model and proprietary models at 12 points before this release and expected about 3 afterward. Its agentic results moved most: Terminal-Bench 4.0 at 67 percent, against 13 percent for K3, and GDPval-AA, a measure of work-style tasks, rose 312 points to an Elo of 1,846. Gains elsewhere were smaller. Humanity’s Last Exam went from 47 percent to 52 percent, and the omniscience score climbed from 20 to 30.
Price is where the picture gets less tidy. The API rates are $2 per million input tokens and $10 per million output tokens, a third lower than K3, but the model used 55,000 output tokens per index task against 48,000 for K3, so the cost per task stayed near $1.98. Hallucination is the stated weakness: the evaluator measured a 44 percent rate on its omniscience test, improved from 53 percent.
A deliberate cap on cyber capability
Moonshot says it removed cyber-related data from two training stages and aims not to release open weights that would extend the absolute frontier of cyber capability. On its internal exploit benchmark the success rate fell from 90.5 percent to 81 percent, still far above K3’s 33.3 percent. The company describes this as a policy it plans to loosen gradually, and it says the model stays behind proprietary models that have been public for at least three months.
That makes K3.1 Pro an unusual case: the weights are permanent once published, so the cap is a decision about every future user, including those who would fine-tune the restriction away. Whether training-stage data removal survives fine-tuning is not addressed in the sources reviewed here.
Where the rest of the field stands
K3.1 Pro leads BenchLM’s open-weight ranking in November at 75 out of 100, ahead of Kimi K3 at 69.6, and 34 of its top 100 overall models are open weight. The count of 2026 releases favors open models, 223 against 171 proprietary, though BenchLM notes those counts only include models that publish benchmark results. Mistral published weights for Mistral Large 4 at the end of October after an API preview, according to a roundup by Flavio Copes, which x9m8 covered in the earlier piece on its trillion-parameter sparse design. Alibaba, by contrast, has named four Qwen4 variants without confirming that weights exist for any of them as of November 6, according to Tech Insider.
For readers new to how these networks work, the earlier explainer on how large language models like ChatGPT work covers the mixture-of-experts idea from the ground up.
Sources
- Kimi K3.1 Pro model card, Hugging Face
- Kimi K3.1-Pro License, Hugging Face
- TechNode, November 5, 2026
- Artificial Analysis, Kimi K3.1 Pro launch benchmarks, October 23, 2026
- VKTR on the K3.1 Pro benchmarks and cyber limits
- BenchLM, open-weight models explained
- Open Source Initiative, Open Source AI Definition 1.0
- Build Fast with AI, Kimi K3 review, July 17, 2026
- Analytics Vidhya, Kimi K3.1 Pro review, October 25, 2026
- Flavio Copes, open-weight AI models
- BuildMVPFast on K3.1 Pro pricing and outside testing
- Tech Insider on Qwen4 weights, November 2026
- WorldofAI on YouTube
