Kimi K3.1 Pro's open weights: 1.56 TB, MXFP4, and a $20M revenue license line

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.

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.

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.

A mixture-of-experts model holds all its experts in memory but uses only a few for each token. So two numbers matter.

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 file is a custom "Kimi K3.1 Pro License." It starts from MIT-style wording, free of charge and "without restriction," then adds conditions.

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.

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.