Moonshot AI releases 2.8T-parameter Kimi K3.1 Pro weights on Hugging Face
Moonshot AI released the model weights of the 2.8T-parameter Kimi K3.1 Pro on Hugging Face, trained and released in 4-bit precision to avoid the 5.6TB storage required for a 16-bit version. In official tests, the model scored 65.7% on Terminal-Bench 4.0, close to Claude Opus 5.5's 66.4%, but without tools it scored only 39.4% on HLE-Diamond, below GPT-6 Astra's 65.8%. The model activates only 104B parameters per token, and 69 of its 93 layers use linear attention to reduce the processing cost of a 1M-token context; a modified MIT license permits internal use, but model-hosting companies with annual revenue over $20 million must negotiate separately.
Why read this
SelectedAI score 88
The original text provides the parameter scale, quantization precision, benchmark comparisons, and license restrictions, allowing readers to judge the openness boundaries of open-source large models based on this information.
Topics
View original postReleasing the model weights of Kimi K3.1 Pro: https://t.co/tJXcYUeSCj We'll keep advancing open intelligence. https://t.co/BbWCJUYYSQ
View original
Releasing the model weights of Kimi K3.1 Pro: https://t.co/tJXcYUeSCj We'll keep advancing open intelligence. https://t.co/BbWCJUYYSQ
Body
Moonshot AI released the massive 2.8T-parameter Kimi K3.1 Pro weights on Huggingface.
On Terminal-Bench, its head-to-head with Claude 5.5
> Trained with 4-bit weights, it also ships at that precision, avoiding the 5.6TB a 16-bit copy would need.
> Moonshot traced steadily rising exploit skill to cyber content buried in its general mid-training corpus and RL environments, and removed both before opening the weights.
> In Moonshot’s own tests the model scores 65.7% on Terminal-Bench 4.0, near Claude Opus 5.5’s 66.4%, but without tools reaches only 39.4% on HLE-Diamond, against GPT-6 Astra’s 65.8%.
> Only 104B of its parameters fire per token, and 69 of its 93 layers use linear attention, which makes a 1M-token context cheaper to process.
> Its modified MIT license exempts internal use but makes model-hosting firms earning over $20M a year negotiate, reining in resellers that could undercut the company’s own API.