Cursor has confirmed that its new Composer 2 coding model was built on top of Moonshot AI’s Kimi K2.5 after users spotted references to Kimi and pressed the company for an explanation.
Cursor VP of developer education Lee Robinson said only about a quarter of the compute behind the final model came from the base model, with the rest coming from Cursor’s own training.
Cursor left Kimi out of the launch
TechCrunch reported that Cursor executives acknowledged that Composer 2 started with Moonshot AI’s open-source Kimi K2.5 model, before Cursor added more training on top of it. Moonshot’s Kimi account backed Cursor’s explanation and said that the integration occurred through Fireworks AI as part of an authorized commercial partnership.
Cursor’s own launch materials presented Composer 2 as a new model from Cursor, without naming the outside base model that helped get it there. According to TechCrunch, Cursor co-founder Aman Sanger said, “It was a miss to not mention the Kimi base in our blog from the start.”
What enterprises can actually take from this
Cursor’s disclosure does not change the product’s published price or its coding focus, but it does answer the question users were asking after launch: what model was under Composer 2 before Cursor added its own work.
Cursor’s March 19 launch post for Composer 2 described the model as “frontier-level” for coding and priced it at $0.50 per million input tokens and $2.50 per million output tokens, but it did not identify Kimi as the starting point.
The same post said Composer 2 scored 61.7 on Terminal-Bench 2.0 and 73.7 on SWE-bench Multilingual, up from 47.9 and 65.9 for Composer 1.5. Cursor also said those gains came from its first continued pretraining run, which it used as a stronger base for reinforcement learning. Cursor also introduced a faster variant priced at $1.50 per million input tokens and $7.50 per million output tokens.
The launch post presented Composer 2 around coding performance, benchmark gains, and pricing, while leaving out the base model behind it. For teams comparing AI coding tools, that missing detail can matter when they are checking how a vendor describes a model’s origin, training path, and commercial partnerships.
Also read: Anthropic’s new institute shows how AI governance and transparency are moving closer to the product conversation.