Moonshot AI released Kimi-K3.1-Pro on Hugging Face
Model: Kimi K3 · availability, license and releases
Moonshot AI released Kimi-K3.1-Pro, an open-weight multimodal agentic model with 2.8T total parameters and a 1M-token context window, on Hugging Face. The model is post-trained from Kimi K3 with joint reinforcement learning across hundreds of domains. It is available under a custom license.
China context
- Original name
- 月之暗面
- Outside China
- Open weights · huggingface.co
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Builders outside China can download the weights from Hugging Face, but must review the custom license before commercial use and run their own benchmarks since no independent evaluations are provided.
- For investors
- Investors should verify whether the custom license allows commercial deployment and whether Moonshot AI's open-weight strategy affects its API business or competitive positioning.
Moonshot AI published Kimi-K3.1-Pro on Hugging Face, describing it as an open-weight, natively multimodal agentic model and the first in the K3.1 series. The model card lists 2.8T total parameters, 104B activated parameters, a 1,048,576-token context length, and a Mixture-of-Experts architecture with 896 experts. It is post-trained from Kimi K3 using large-scale joint reinforcement learning across hundreds of downstream domains, targeting improvements in coding, 3D creation, financial research, factual accuracy, and other agentic capabilities. The repository includes links to a tech blog and API quickstart, and the license is listed as 'other'.
Kimi-K3.1-Pro uses a Mixture-of-Experts architecture with 2.8T total parameters and 104B activated parameters, 93 layers, 96 attention heads, and a 1M-token context window. The attention layer composition includes 69 KDA layers and 24 Gated DSA layers, with a DSA index top-K of 2048. The model card states it is post-trained from Kimi K3 with large-scale joint reinforcement learning across hundreds of downstream domains, but no independent benchmark results are provided in the evidence.
Developers outside China can now access a 2.8T-parameter open-weight multimodal model from Moonshot AI, but the custom license and lack of independent benchmarks mean adoption requires legal review and internal evaluation before deployment.
The release gives enterprises and developers a large open-weight multimodal model option, but the custom license and absence of independent performance data create uncertainty about cost, compliance, and capability relative to alternatives.
The next verifiable signals are whether Moonshot AI publishes benchmark results or third-party evaluations, and whether the custom license permits commercial use outside China. The linked tech blog and API quickstart may provide additional details on intended use and deployment.