Moonshot AI Releases Kimi K3.1 Pro Weights

Moonshot AI released Kimi K3.1 Pro’s weights through Hugging Face on the 4th, local time. Developers can run and modify the model on their own infrastructure. The release followed the schedule announced when the model launched last month. The repository occupies about 1.56 terabytes (TB), with the weights divided into 96 files; model configuration and input-processing code are also included.

Kimi K3.1 Pro is a Mixture of Experts (a model that activates selected expert networks) with 2.8 trillion parameters, of which 104 billion are activated during inference. It processes text and images and accepts context of up to about 1 million tokens. The model card describes an architecture combining KDA and DeepSeek Sparse Attention (DSA). Of 93 layers, 69 use KDA and 24 use gated DSA. The model replaces the gated MLA layers that performed full attention in the earlier Kimi K3 with DSA. Moonshot AI says DSA selects which input tokens to reference, improving computational efficiency for long-context processing.

The weights are distributed in MXFP4 format, while intermediate values calculated during operation, called activations, use MXFP8. The model also uses Quantization-Aware Training (training that accounts for low-precision calculations from the start). Moonshot AI recommends the vLLM, SGLang, and TokenSpeed inference engines for developers running the model on their own infrastructure. Developers can download the weights and run the model themselves or use Moonshot AI’s API.

The proprietary license allows use, modification, fine-tuning, redistribution, and commercial use. However, model service providers and their affiliates whose combined revenue exceeds $20 million over any consecutive 12-month period must sign a separate agreement before commercial use. Commercial products or services must display the model name if they exceed 100 million monthly active users or $20 million in monthly revenue. Internal use, as well as use through Moonshot AI’s official products and certified inference partners, is exempt from the contract and display requirements.

Kimi K3.1 Pro is designed for agent tasks such as coding, 3D production, and financial research. In its earlier launch announcement, Moonshot AI said it had selectively removed cybersecurity-related content from intermediate training data and reinforcement-learning tasks to reduce potential misuse of the released weights. The company said the measure was intended to lower vulnerability-exploitation capabilities while preserving general coding performance.

Moonshot AI released Kimi K3.1 Pro’s weights through Hugging Face on the 4th, local time. Developers can run and modify the model on their own infrastructure. The release followed the schedule announced when the model launched last month. The repository occupies about 1.56 terabytes (TB), with the weights divided into 96 files; model configuration and input-processing code are also included.

Kimi K3.1 Pro is a Mixture of Experts (a model that activates selected expert networks) with 2.8 trillion parameters, of which 104 billion are activated during inference. It processes text and images and accepts context of up to about 1 million tokens. The model card describes an architecture combining KDA and DeepSeek Sparse Attention (DSA). Of 93 layers, 69 use KDA and 24 use gated DSA. The model replaces the gated MLA layers that performed full attention in the earlier Kimi K3 with DSA. Moonshot AI says DSA selects which input tokens to reference, improving computational efficiency for long-context processing.

The weights are distributed in MXFP4 format, while intermediate values calculated during operation, called activations, use MXFP8. The model also uses Quantization-Aware Training (training that accounts for low-precision calculations from the start). Moonshot AI recommends the vLLM, SGLang, and TokenSpeed inference engines for developers running the model on their own infrastructure. Developers can download the weights and run the model themselves or use Moonshot AI’s API.

The proprietary license allows use, modification, fine-tuning, redistribution, and commercial use. However, model service providers and their affiliates whose combined revenue exceeds $20 million over any consecutive 12-month period must sign a separate agreement before commercial use. Commercial products or services must display the model name if they exceed 100 million monthly active users or $20 million in monthly revenue. Internal use, as well as use through Moonshot AI’s official products and certified inference partners, is exempt from the contract and display requirements.

Kimi K3.1 Pro is designed for agent tasks such as coding, 3D production, and financial research. In its earlier launch announcement, Moonshot AI said it had selectively removed cybersecurity-related content from intermediate training data and reinforcement-learning tasks to reduce potential misuse of the released weights. The company said the measure was intended to lower vulnerability-exploitation capabilities while preserving general coding performance.