Cursor's New AI Model May Be Built on Kimi K2.5

Cursor released Composer 2 yesterday, positioning it as a self-developed intelligent agent model designed specifically for advanced coding work in its editor. The company stated that the model achieved excellent benchmark results in complex tasks, long-term planning, tool use, and multi-step reasoning, and is sold at 50 cents per million input tags, with a faster variant enabled by default. The market response has been enthusiastic. Developers have subsequently begun to delve deeper into its development.

Within hours of launch, a tester spotted an unusual model ID "kimi-k2p5-rl-0317-s515-fast" from API calls. The ID points directly to Kimi K2.5, a model developed by Chinese AI lab Moonshot AI, with reinforcement learning applied on top.

Moonshot's pre-training leader went further(the tweet was deleted). He examined the tokenizer fingerprints and confirmed they matched Kimi K2.5 exactly. Moonshot has since stated that Cursor did not obtain a license or pay any fees to use the model.

Despite Kimi K2.5 being released under a modified MIT license, the license includes a specific condition for large commercial products. It requires clear and prominent credit to those with significant revenue or a sizeable user base. Cursor's announcement and official documentation mention no such origin.

Cursor, for its part, describes Composer 2 as an in-house model trained with continued pre-training and reinforcement learning for better performance in agentic coding setups. That framing says nothing about a foundation built on someone else's work.

To be fair, the model itself performs. Kimi K2.5 supports up to 262,000 tokens of context and is well-suited for agent workflows, code edits, and terminal operations. If Cursor fine-tuned it effectively, the result is still a capable tool delivered at a competitive price. Users get something genuinely useful.

But usefulness and honesty are separate questions. The developer community has taken notice, and the debate is growing. Building on open models is common practice. Skipping attribution when the license explicitly requires it is a different matter, especially for a company operating at Cursor's scale.

Cursor has not publicly responded to the claims. Moonshot remains firm. The community is waiting. Meanwhile, Composer 2 keeps running. It is fast, affordable, and functional. Whether the story behind it holds up is still an open question.

Cursor and Moonshot AI did not respond to requests for comment at the time of publication.