Illustration, AI-generated from the story text. Tap a panel to read the story through it.

1 / 4Moonshot publishes the full 2.8-trillion-parameter Kimi K3.1 Pro weights by the date it promised×

The release

The weights are on Hugging Face with no access gate: 96 shards, about 1.56 terabytes, in MXFP4 format, covering a mixture-of-experts model with 104 billion active parameters and a context window of about 1 million tokens. Reports of the release began circulating by about 16:30 UTC on Wednesday, the last day of Moonshot's promise, though the repository itself carries a November 2nd creation date.

The license

The license is custom. It opens like MIT and allows fine-tuning and resale, then requires a separate agreement from any company selling model access whose revenue tops $20 million over 12 months, and a visible "Kimi K3.1-Pro" credit on products above 100 million monthly users or $20 million in monthly revenue. So the largest inference resellers have to negotiate with Moonshot and almost everyone else does not.

The score

On the Vals Index it scores 62.52%, up from 52.96% for Kimi K3 and just ahead of DeepSeek V4.1 Flash at 61.26%, though it fell on legal research and tax agent tests.

What it takes

Planning figures from the parent model point to at least one eight-GPU server of the newest class. Tencent's final Hy4 weights were still missing from Hugging Face this afternoon.

Sources and the full story

‹›

Moonshot publishes the full 2.8-trillion-parameter Kimi K3.1 Pro weights by the date it promised

The weights are on Hugging Face with no access gate: 96 shards, about 1.56 terabytes, in MXFP4 format, covering a mixture-of-experts model with 104 billion active parameters and a context window of about 1 million tokens. Reports of the release began circulating by about 16:30 UTC on Wednesday, the last day of Moonshot's promise, though the repository itself carries a November 2nd creation date.

The license is custom. It opens like MIT and allows fine-tuning and resale, then requires a separate agreement from any company selling model access whose revenue tops $20 million over 12 months, and a visible "Kimi K3.1-Pro" credit on products above 100 million monthly users or $20 million in monthly revenue. So the largest inference resellers have to negotiate with Moonshot and almost everyone else does not.

On the Vals Index it scores 62.52%, up from 52.96% for Kimi K3 and just ahead of DeepSeek V4.1 Flash at 61.26%, though it fell on legal research and tax agent tests. Planning figures from the parent model point to at least one eight-GPU server of the newest class. Tencent's final Hy4 weights were still missing from Hugging Face this afternoon. SourcesACC

Sources

• Amoonshotai/Kimi-K3.1-Pro (Hugging Face; API, model card and license read directly)huggingface.co
• CMoonshot makes Kimi K3.1 Pro weights free for internal use (BERI)beri.net
• CMoonshot AI open-sources Kimi K3.1 Pro weights (PANews)panews.io

This story ran in the Afternoon Brief of November 4th 2026.