Kimi K3.1 Pro: what its released weights mean
Moonshot releases Kimi K3.1 Pro weights: architecture, deployment requirements and licence conditions to review.
Checked against Moonshot’s model card and announcement; licence terms apply.
Kimi K3.1 Pro weights are available in Moonshot AI’s Hugging Face repository. The release allows teams to evaluate and deploy the model on their own infrastructure, subject to the licence. It follows the hosted-service launch announced in October.
Architecture and size
The model card describes a Mixture-of-Experts architecture with 2.8 trillion total parameters, around 104 billion active parameters and a context window of roughly one million tokens. The MXFP4 weights are large: plan storage, memory and server compute before downloading.
Licence and deployment
Open weights do not mean unconditional use. Review the full Kimi K3.1 Pro licence, especially provisions for model services and large-scale products, before commercial deployment. A hosted API may be more practical without sufficient infrastructure.
Evaluation
Moonshot reports benchmark results for coding and agent tasks. These are vendor figures; compare the model on your own examples, actual costs, latency and human review before integrating it into production.