On March 19, Cursor announced Composer 2 as its most capable proprietary coding model, built with "continued pretraining" and "scaled reinforcement learning." Less than 24 hours later, developer Fynn (@fynnso) spotted a detail Cursor hadn't disclosed: the internal model ID exposed through the API reads accounts/anysphere/models/kimi-k2p5-rl-0317-s515-fast.

That's not a Cursor model name. That's Kimi K2.5 โ€” the open-weight model from Chinese AI lab Moonshot AI โ€” with RL fine-tuning applied on top.

What the License Says

Kimi K2.5 ships under a Modified MIT License with one significant clause: any commercial product earning more than $20 million per month must "prominently display 'Kimi K2.5' on the user interface." Cursor's annualized revenue is reported at approximately $2 billion โ€” roughly 8ร— over that monthly threshold. The Composer 2 announcement mentioned Kimi K2.5 nowhere.

Moonshot's Response

Yulun Du, Moonshot AI's head of pretraining, publicly confirmed the tokenizer match between Composer 2 and Kimi K2.5. Two other Moonshot employees also confirmed Cursor wasn't properly licensed, then deleted their posts.

Cursor, valued at $29.3 billion, has not issued a public response. The story is trending on Hacker News, where commenters noted that Cursor Composer 1 was reportedly built on Qwen โ€” suggesting a pattern of shipping commercial products on open-weight Chinese models without clear attribution. Moonshot AI has not yet filed a formal legal claim.