Moonshot Releases Full Kimi K3.1 Pro Weights: 2.8T Parameters, About 1.56 TB

What happened

The weights went live on Nov 4 in the Hugging Face repository moonshotai/Kimi-K3.1-Pro, and Moonshot's official X account confirmed the release the same day at 16:00 UTC with the post "Releasing the model weights of Kimi K3.1 Pro". The repository holds 96 files totaling about 1.56 TB. The model card describes it as the first model in the K3.1 series and the company's most capable 3T-class model to date; at its Oct 24 launch it shipped first as an API and in Moonshot's apps.

The model card lists the full specifications: 2.8 trillion total parameters with 104 billion activated per token; a mixture-of-experts design with 896 experts, 16 selected per token; 93 layers and a 1,048,576-token context; MXFP4 weights with MXFP8 activations, trained with quantization awareness; and the MoonViT-V2 vision encoder at 401 million parameters, with text and image input.

The license and deployment terms were published together. The weights use Moonshot's own Kimi K3.1 Pro License, MIT-style: use, copying, modification, fine-tuning and sale are allowed, with the copyright notice retained. Operators of a model-as-a-service business whose combined annual revenue exceeds $20 million must sign a separate agreement with Moonshot before commercial use; products with more than 100 million monthly active users or more than $20 million in monthly revenue must display "Kimi K3.1-Pro" prominently in their interface. Internal use, and use through official products or certified inference partners, is exempt from both conditions. The model card recommends the vLLM, SGLang and TokenSpeed inference engines, and the API model name kimi-k3.1-pro stays available.

Key facts

1. Publisher and date: Moonshot AI, Nov 4, 2026; the official X account confirmed the release the same day at 16:00 UTC
2. Repository: Hugging Face moonshotai/Kimi-K3.1-Pro, 96 files totaling about 1.56 TB
3. Scale: 2.8 trillion total parameters, 104 billion activated per token; 896 experts, 16 selected per token
4. Context: 1,048,576 tokens (1 million tokens)
5. Quantization: MXFP4 weights and MXFP8 activations, trained with quantization awareness
6. Vision: MoonViT-V2 encoder at 401 million parameters; text and image input
7. Benchmarks (company-reported): 65.7% on Terminal-Bench 4.0, against 66.4% for Claude Opus 5.5 and 58.2% for GPT-6 Astra in the same model-card table
8. License: Kimi K3.1 Pro License, MIT-style; model-as-a-service operators with over $20 million in annual revenue must sign a separate agreement
9. Deployment: the model card recommends vLLM, SGLang and TokenSpeed; the API model name is kimi-k3.1-pro
10. Promise kept: the Oct 24 launch promised the weights by Nov 4, and they landed on Nov 4