Kimchi Pro belt and rode

 

Chinese AI makes another major move: Kimi K3.1 Pro open-sources its full 1.56TB weights, allowing commercial use.

50 minutes ago

Beating AI News (Dongcha) reports that Chinese AI startup Moonshot AI has officially open-sourced the full model weights of its Kimi K3.1 Pro. The model was already rolled out for applications and APIs in late October; the released Hugging Face repository includes 96 weight shards with a total size of approximately 1.56 TB. Kimi K3.1 Pro inherits the custom licensing terms of its predecessor K3, permitting free use, modification, distribution, and commercial use, but with two special restrictions: 1. Large API service providers must sign a separate agreement: If an enterprise runs model inference or fine-tuning services, and its own and affiliated entities have combined total revenue exceeding $20 million over 12 consecutive months, it must sign an additional agreement with Moonshot AI prior to commercial use. 2. Ultra-large commercial products must label the model name: If a product using the model has over 100 million monthly active users (MAUs) or monthly revenue exceeding $20 million, it must prominently display "Kimi K3.1-Pro" on its interface. Pure internal use, or calls to the model via Moonshot AI’s official products and certified inference partners, are exempt from the above signing and labeling requirements. To mitigate cyberattack risks stemming from open weights, Moonshot AI removed cyberattack-related content from mid-training data and reinforcement learning tasks, actively restricting the model’s vulnerability exploitation capabilities. Official internal tests show the unmodified model (without these restrictions) achieved a 90.5% pass rate in vulnerability exploitation tasks after up to three attempts; the final released version cut this to 81.0%, which remains significantly higher than predecessor K3’s 33.3%.