Cursor launched Composer 2 last week with a new model that benchmarks at GPT-5 tier for coding tasks, at roughly 10x lower cost than Claude Opus. Impressive numbers. The catch: the model is built on top of Kimi K2.5, an open-weight model from Moonshot AI, and Cursor didn’t mention that publicly.
The connection surfaced when people found the Kimi model name embedded in Cursor’s API URLs. Moonshot AI then posted a congratulatory tweet confirming it: “We are proud to see Kimi-k2.5 provide the foundation.”
To be clear, Cursor did actual work here. They applied roughly 3x additional compute in post-training and reinforcement learning on top of the base Kimi K2.5 weights. That’s not nothing. The resulting model apparently performs well enough to justify the benchmarks they’re claiming.
The licensing situation is also technically clean. Kimi K2.5 uses a modified MIT license that requires prominent attribution for large commercial products. Cursor routed inference through Fireworks AI as a partner, which appears to satisfy the license terms on paper.
But “technically compliant” and “transparent” aren’t the same thing. Cursor presented Composer 2 as their model without mentioning it was built on an open-weight base from a Chinese AI lab. Users reasonably assumed they were getting something Cursor built from scratch, or at least a fine-tune of a well-known model like Claude or GPT-4.
The broader concern is what this does to incentives. Smaller labs that release open-weight models are betting that the community goodwill and attribution they get back is worth more than keeping the weights closed. If companies can quietly build on those weights, ship a product under their own branding, and only acknowledge the source when caught, that bet starts looking worse. Future labs may just keep their models closed.
Cursor’s technical achievement here is real. The transparency problem is also real. Both things are true.

Collections

Cursor launched Composer 2 claiming GPT-5-tier coding performance at 10x lower cost than Claude Opus, but didn't disclose that the model is built on Kimi K2.5, an open-weight model from Moonshot AI. The connection was discovered when users found the Kimi model name in Cursor's API URLs, later confirmed by Moonshot AI. Cursor did apply significant post-training compute on top of the base weights, and the licensing appears technically compliant via a modified MIT license. However, the lack of transparency raises concerns about incentives for open-weight model releases — if companies can quietly build on open weights and only credit the source when caught, future labs may opt to keep models closed.

Kimi-k2.5 powers Cursor's Composer 2