Moonshot's Kimi K3.1 Pro goes open-weight: 2.8T parameters, a 1M-token window and a license with a $20M catch

Moonshot AI has put Kimi K3.1 Pro on Hugging Face. The model card lists 2.8 trillion total parameters, 104 billion active per token, a context length of 1,048,576 tokens (about 1 million) and text and image input. Those match the figures Moonshot gave at launch. The card does not print a release date, and we could not independently pin down the day the files went up. Moonshot had said the weights would arrive by November 4.

The model was first sold through Moonshot's own apps and API on October 23, according to VKTR, which described it as post-trained from Kimi K3, a 2.8-trillion-parameter model introduced in July. An agency report in the Qatar Tribune carried the same launch and the same weights-by-November-4 promise. For a business, the question is not whether the model is clever. It is what open weights change.

The benchmarks: strong in places, thin in others

Every score below is reported by Moonshot, not independently verified, and several tests on the card are marked internal. The comparison table, as relayed by VKTR, sets K3.1 Pro against Claude Opus 5.5 and GPT-6 Astra.

Benchmark (vendor-reported)	Kimi K3.1 Pro	Claude Opus 5.5	GPT-6 Astra
Terminal-Bench 4.0	65.7	66.4	58.2
SWE Marathon v1.1	51.7	46.3	42.5
OfficeQA Pro v2	64.4	62.2	54.4
Program Bench	31.5	65.0	50.0
Terminal-Bench-Science 0.1	38.1	63.3	68.1

Read it straight: Moonshot's own table shows the model ahead on a long software-engineering task suite and an office-documents test, level on terminal work, and far behind on two others. The model card also lists GPQA-Diamond at 94.2, BrowseComp at 92.1 and Humanity's Last Exam at 51.1 without tools. Moonshot says its hallucination rate on an internal adversarial test fell from 23.4% to 8.3%, and itself cautions that the test is not representative of everyday error rates. LetsDataScience, citing developer Simon Willison, reports the model near the top of the Artificial Analysis index, and notes that this does not validate each of Moonshot's own numbers. Test it on your own work before believing any table.

What "open weights" gives a business

Control of the data path. With the weights you can run the model inside your own cloud account or data center. Prompts, contracts, customer records and code never go to a third-party API. For a law firm, a clinic group or a lender with confidentiality duties, that is the real feature, more than any benchmark.

Price you set yourself. Moonshot's hosted API is listed at $2 per million input tokens and $10 per million output tokens, as reported by LetsDataScience. A team sending 100 million input tokens and 20 million output tokens a month would pay about $200 plus $200, or $400, on that rate. Self-hosting only wins when you keep very expensive hardware busy. At low volume the API is cheaper.

No forced upgrades. A model you host does not change under you. Pin the version and your prompts and tests keep behaving the same way, which matters for anything audited.

Tuning. You can fine-tune on your own documents, subject to the license below.

The cost nobody puts in the headline: the hardware

The card says the weights use MXFP4 with MXFP8 activations, trained with quantization in the loop, and recommends the vLLM, SGLang and TokenSpeed inference engines. It lists no hardware requirements. Our own rough arithmetic: 2.8 trillion parameters at about 4 bits each is roughly 1.4 terabytes for the weights alone, before the memory a 1-million-token context needs. That is a multi-GPU server cluster, not a workstation. Only 104 billion parameters are active per token, which helps speed, but the full set must still sit in memory. Most small and mid-size firms will reach the model through a cloud host or Moonshot's API, and open weights still matter to them because competing hosts can offer it, which keeps prices honest.

The license: read Sections 2, 3 and 4

The card names the "Kimi K3.1 Pro License" and links the LICENSE file. As we read it, it is not a plain permissive license:

• A company running a "Model as a Service" business with more than $20 million in revenue over any 12 months must sign a separate agreement with Moonshot before commercial use.
• A product with over 100 million monthly active users, or over $20 million in monthly revenue, must show "Kimi K3.1-Pro" in its interface.
• Internal use that does not expose the model or its outputs to third parties is exempt, as is use through Moonshot's official products or certified inference partners.

For nearly every small business the practical reading is that internal use and ordinary products built on the model are clear. A firm reselling model access as its core product should have counsel read the full text, since we have summarized it, not interpreted it for your case.

What Moonshot held back

Moonshot says it removed cyber-related content from training and cyber tasks from its reinforcement-learning environments, and that it will not release open weights that "would extend the absolute frontier of cyber capability," per VKTR. On its internal exploit suite it reports the released model solved 81% of problems, against 90.5% for an unmitigated checkpoint. That is Moonshot's claim on Moonshot's test. If you plan to use the model for security work, expect a weaker tool than the closed-lab alternatives.

What to do

If you handle regulated or confidential data and already use AI tools, run a two-week bake-off: your own 50 real tasks, the open model against your current one, scored by the people who do the work. If you do not, the sensible move is to wait for hosted versions and see what independent testers publish. Related: our coverage of China's AI funding race, in the DeepSeek funding round story.

Sources: Hugging Face model card and LICENSE (Moonshot AI); VKTR; Qatar Tribune; LetsDataScience. All benchmark figures are vendor-reported. The memory estimate is the Chronicle's own arithmetic. This is information, not investment or legal advice.