Cursor Admits Composer 2 is Based on Moonshot's Kimi K2.5

The Cursor launched its new AI programming model, Composer 2, on Wednesday (19), touted as frontier-level and "in-house." Less than 24 hours later, the company admitted that the model is based on the Kimi K2.5 from Chinese company Moonshot AI.

Everything started when a developer intercepted the API call and discovered the exact model ID: kimi-k2p5-rl-0317-s515-fast. The name reveals it: it's the Kimi K2.5 with reinforcement learning (RL) on top. Yulun Du, head of pretraining at Moonshot, tested the tokenizer and confirmed it was identical to the original model. He publicly questioned Cursor about the license compliance — posts that were deleted within hours.

In response, Lee Robinson, VP of Developer Education at Cursor, openly confirmed: "KIMI K2.5. Yes, that's the base we used." According to him, only about 1/4 of the compute was from the base model, with the rest being their own training. The company claims to be following the license through terms with inference partners like Fireworks.

The issue lies in the Modified MIT license of Kimi K2.5. It requires commercial products grossing over $20 million per month or having over 100 million users to prominently display "Kimi K2.5" in the interface. Cursor, which had listed the model as free weeks ago, now uses the relabeled version to justify a potential $50 billion valuation — up from $29.3 billion from the previous round.

The Kimi K2.5 is a 1 trillion parameter model trained with 15 trillion tokens. Moonshot made it open to foster the ecosystem. Meanwhile, discussions are blowing up on Hacker News and Reddit, with many questioning whether Cursor's strategy — a VS Code fork packaging open-source models — justifies billion-dollar valuations.

The case is expected to pressure other startups. Industry analysts predict Chinese companies will adopt more restrictive licenses going forward, and investors will demand more transparency about the real bases of "in-house" models.