Moonshot put the full Kimi K3.1 Pro weights on Hugging Face, as it promised for November 4. We measured the download, did the memory math for its 2.8 trillion parameters, and read the custom license clause by clause.
This explains reporting by
Moonshot AI, Kimi K3.1 Pro model card on Hugging Face (repo created November 2, 2026). Read the original first:
https://huggingface.co/moonshotai/Kimi-K3.1-Pro
In one minute #
- Kimi K3.1 Pro is a mixture-of-experts model: 2.8 trillion parameters in total, 104 billion used for each token, about 1 million tokens of context, text and images.
- The weights take 1.56 TB on Hugging Face, in 96 safetensors files. That is about 4.5 bits per parameter on average, because the experts are stored in MXFP4.
- Moonshot reports 65.7% on Terminal-Bench 4.0, next to 66.4% for Claude Opus 5.5 and 58.2% for GPT-6 Astra. These are Moonshot's own numbers.
- The license is not standard open source. A company with over $20 million in revenue over 12 months needs a separate deal before selling it as a service. Internal use is exempt.
- You need well over 1.5 TB of fast memory just to hold the weights, so this is a multi-GPU server model, not a laptop model.
What was released #
Moonshot's blog said in October that K3.1 Pro was live in its own apps and API and that "the full model weights will be released by Nov 4." The Hugging Face repo moonshotai/Kimi-K3.1-Pro is now public. The Hugging Face API shows it was created on November 2, 2026.
The model card describes the design:
- 2.8T total parameters, 104B activated per token.
- 93 layers, 1 of them dense. 896 experts, 16 picked per token, plus 2 shared experts.
- Attention: 69 KDA layers and 24 Gated DeepSeek Sparse Attention layers.
- A 401M-parameter vision encoder (MoonViT-V2), so it reads images as well as text.
- MXFP4 weights with MXFP8 activations, trained with quantization in the loop.
- Context of 1,048,576 tokens.
- Suggested engines: vLLM, SGLang or TokenSpeed.
The download, measured #
We asked the Hugging Face API for the file list and sizes instead of trusting a headline number.
curl -s "https://huggingface.co/api/models/moonshotai/Kimi-K3.1-Pro?blobs=true"
files: 112 (96 are .safetensors)
safetensors bytes: 1,561,293,160,568 (about 1.56 TB)
parameter count: 2,780,110,363,392 (reported by the API)
Divide the bytes by Moonshot's 2.8 trillion parameters and you get about 4.5 bits per parameter. That fits the card: most of the weight sits in the experts in a 4-bit format, while about 57 billion parameters are stored in 16-bit (BF16), per the same API response.
Practical meaning: at 1 gigabit per second, 1.56 TB takes about three and a half hours to download. Plan the disk and the bandwidth before you plan the GPUs.
What 104B active parameters means for serving #
A mixture-of-experts model holds all its experts in memory but uses only a few for each token. So two numbers matter.
- Memory to hold it: all 2.8T parameters, about 1.56 TB, plus room for the context cache and the engine. That is far beyond one GPU. It is a job for a multi-GPU server or several of them.
- Work per token: only 104B parameters. At the same 4.5 bits each, that is roughly 58 GB of weights read for every token in a single stream.
The second number sets speed. When one user is generating, the GPU mostly waits on memory. A rough ceiling is memory bandwidth divided by about 58 GB. With 8 TB per second of total bandwidth, that is about 140 tokens per second before any overhead. Batching many users helps, because they share the same weight reads.
This is arithmetic, not a benchmark. Real speed depends on how the engine spreads the experts across GPUs and on how much context each request carries.
The license, clause by clause #
The license file is a custom "Kimi K3.1 Pro License." It starts from MIT-style wording, free of charge and "without restriction," then adds conditions.
- Section 2, model as a service: if your company and its affiliates earn more than US$20 million over any 12 months in a row, you "must enter into a separate agreement with Moonshot AI" before commercial use. This section does not cover products where the model is only embedded inside specific features, or simple relaying of requests to models hosted by others.
- Section 3, display: a commercial product with more than 100 million monthly active users, or more than US$20 million in monthly revenue, must show "Kimi K3.1-Pro" prominently in its interface.
- Section 4, exemptions: sections 2 and 3 do not apply to internal use that does not make the model, its outputs or its capabilities available to third parties. They also do not apply when you use it through Moonshot's own products or certified inference partners.
- Section 5: the usual as-is, no warranty, no liability.
For most IT teams and MSPs the reading is: running it inside your own company for your own work is exempt. Selling access to it as a hosted model, once you are past $20 million in revenue, needs a deal first.
How good is it #
Moonshot's blog puts it at 65.7% on Terminal-Bench 4.0, "narrowly behind Claude Opus 5.5 (66.4%)" and ahead of GPT-6 Astra at 58.2%. The model card lists dozens more scores, including GPQA-Diamond 94.2 and OSWorld V2.1 79.6. Several are Moonshot's internal tests.
One outside data point: Simon Willison asked K3.1 Pro, through OpenRouter, which game a small neural network was trained on from its raw weights. It answered Breakout correctly in 9 minutes 41 seconds, using 19,493 reasoning tokens.
Who is affected #
| Case | Status |
|---|---|
| Teams that want a frontier-class model on their own hardware | Weights are out; needs over 1.5 TB of GPU memory for weights alone. |
| Internal use inside one company | Exempt from the license's revenue and display sections. |
| Hosting it as a paid model service | Separate Moonshot agreement needed above $20M revenue over 12 months. |
| Very large consumer products | Must display "Kimi K3.1-Pro" above 100M monthly users or $20M monthly revenue. |
What to do #
- Read the LICENSE file before any commercial use and check your 12-month revenue against the $20M line.
- Budget storage and bandwidth for a 1.56 TB download.
- Try it first through the hosted API to see if it beats your current model on your own tasks.
- Test it on your workload instead of trusting the self-reported scores.
What is still unknown #
- Real serving speed. Moonshot gives no hardware guidance, and our per-token figure is arithmetic, not a test.
- How much memory the 1M-token context needs. The card does not give the cache size per token.
- Independent benchmark results. The scores on the card and blog are Moonshot's own, and some are internal tests.
- API pricing. The blog links to a pricing page but lists no numbers.
- How "aggregate revenue" is counted for groups with many affiliates. The license does not define it further.
Sources #
AI News Report· every headline, every morning.