NNewsGPT ← Home
Africa

Moonshot AI releases Kimi K3 model weights with new commercial licensing terms

Africa2 hr ago

Moonshot AI has released the weights for its 2.8 trillion parameter Kimi K3 model, totaling 1.56 terabytes on Hugging Face. This follows their previous K2 model, which introduced a modified MIT license in July 2025. The K2 license required attribution for commercial entities exceeding 100 million monthly active users or $20 million in monthly revenue. The new K3 license, which Moonshot AI does not describe as "open source" but rather "open weight," imposes stricter conditions for large-scale commercial use. Specifically, any entity operating a "Model as a Service" business that generates over $20 million in aggregate revenue over any 12-month period must enter into a separate agreement with Moonshot AI before using the K3 model or its derivatives for commercial purposes. OpenRouter is already providing access to K3 through seven different providers, with pricing set at $3 per million input tokens and $15 per million output tokens, matching Moonshot AI's own rates.

AI Analysis

The release of Kimi K3's substantial 2.8 trillion parameter weights, coupled with its evolving licensing structure, highlights a critical tension in the AI development landscape. While "open weight" models foster innovation and broader adoption, the commercial clauses, particularly the revenue thresholds for "Model as a Service" businesses, indicate a strategic move by Moonshot AI to monetize its significant investment. This approach balances the benefits of community engagement and research with the need for sustainable revenue, especially for entities operating at a massive scale. The tiered licensing reflects a growing trend of differentiating access and commercial rights based on the economic impact of AI model usage, prompting developers and businesses to carefully assess their operational scale against these terms to ensure compliance and manage costs in the burgeoning AI economy.

AI-generated to prompt reflection — not editorial opinion, not advice, not a statement of fact. How this works.

Compiled by NewsGPT from Simon Willison. Read the original for full details.