Moonshot AI posts Kimi K3.1 Pro weights on Hugging Face
The 2.8-trillion-parameter mixture-of-experts model, distributed in 96 files, allows commercial use under Moonshot AI's own license with conditions.
Moonshot AI released the weights of its Kimi K3.1 Pro model on Hugging Face on November 4, giving developers the files needed to run and modify the model on their own infrastructure.
The distribution repository is about 1.56 TB and splits the weights across 96 files. Model configuration and input-processing code are distributed alongside the weights.
Kimi K3.1 Pro is a mixture-of-experts model with 2.8 trillion total parameters, activating 104 billion at inference. It handles text and images and accepts context of up to about 1 million tokens. Of its 93 layers, 69 use KDA and 24 use gated DeepSeek Sparse Attention (DSA); the gated MLA layers that performed full attention in Kimi K3 were replaced with DSA.
Weights are distributed in MXFP4 format, with activations in MXFP8. Moonshot AI applied quantization-aware training from the training stage. The company lists vLLM, SGLang and TokenSpeed as recommended inference engines.
License terms
Use, modification, fine-tuning, redistribution and commercial use are permitted under Moonshot AI's own license. Model service providers and affiliates with combined revenue above $20 million over 12 consecutive months must sign a separate contract before commercial use. Commercial products or services exceeding 100 million monthly active users or $20 million in monthly revenue must display the model name on the user interface. Internal use and use via official Moonshot AI products or certified inference partners are exempt from the separate contract and model-name display conditions.
Kimi K3.1 Pro was launched in October 2025, and the weight release was announced at that time. In its launch announcement, Moonshot AI said it filtered out and removed cybersecurity-related content from intermediate training data and reinforcement-learning tasks to reduce misuse risk.
Quick answers
What is Kimi K3.1 Pro?
It is a mixture-of-experts model from Moonshot AI with 2.8 trillion total parameters, activating 104 billion at inference, that handles text and images with context of up to about 1 million tokens.
Can Kimi K3.1 Pro be used commercially?
Yes. Commercial use is permitted under Moonshot AI's own license, but model service providers with combined revenue above $20 million over 12 consecutive months must sign a separate contract, and very large commercial products must display the model name.
Which inference engines does Moonshot AI recommend?
The company lists vLLM, SGLang and TokenSpeed.