Cursor acknowledges its new low-cost coding model has Chinese bones

Cursor just admitted its latest coding model has Chinese roots — after leaving it out at launch. It's also priced far cheaper than rivals. Andria Lo/Reuters 2026-03-23T04:35:35.865Z Share Copy link Email Facebook WhatsApp X LinkedIn Bluesky Threads lighning bolt icon An icon in the shape of a lightning bolt.

Impact Link Save Saved Read in app This story is available exclusively to Business Insider subscribers. Become an Insider and start reading now. Have an account? Log in. Cursor left out one key detail about its new coding model: it started from Kimi K2.5. Composer 2 is cheaper, more capable — and built on a Chinese open-source model, Cursor's executives said. An X user spotted code suggesting Kimi under the hood, sparking disclosure. AI-generated summary Summaries are generated by an AI model trained on Business Insider's articles. AI may make mistakes or provide inaccurate/incomplete information. We're unable to load that answer right now. Please try again. What is Kimi K2.5? How does Composer 2 compare to rivals? How is reinforcement learning used here? Why did Cursor omit Kimi's involvement? What are the benefits of open-source models? Cursor just acknowledged that its latest coding model has Chinese roots — a detail it left out the first time around. Loading audio narration... In a series of posts on X over the weekend, Cursor executives said Composer 2 was initially built on top of Kimi K2.5, an open-source model developed by Chinese startup Moonshot AI."We've evaluated a lot of base models on perplexity-based evals and Kimi k2.5 proved to be the strongest!" said Cursor's cofounder Aman Sanger on X on Saturday. "It was a miss to not mention the Kimi base in our blog from the start," he added.The disclosure appears to have been sparked by an X user named Fynn, who posted on Friday that Composer 2 was "just Kimi 2.5" with additional reinforcement learning. To support the claim, the user pointed to code snippets that appeared to reference Kimi as the underlying system.'At least rename the model ID," the user wrote. In response to the user's X post, Cursor's vice president of developer education, Lee Robinson, acknowledged that Composer 2 was built on Kimi K2.5 as an open-source base."We will do full pretraining in the future," Robinson said. "Only ~1/4 of the compute spent on the final model came from the base, the rest is from our training," he added.Robinson also said the company is complying with the model's licensing terms through its inference provider. The Chinese startup posted on X on Saturday that Cursor is using Kimi K2.5 under an authorized commercial partnership."Seeing our model integrated effectively through Cursor's continued pretraining & high-compute RL training is the open model ecosystem we love to support," the post read. Cursor was last valued at $29.3 billion in November.Cursor's new model is cheaper and betterCursor said in a blog post on Thursday that Composer 2 is "frontier-level at coding" and priced at $0.50 per million input tokens and $2.50 per million output tokens, calling it "a new, optimal combination of intelligence and cost." By comparison, Anthropic's Claude Opus 4.6 is priced at $5 per million input tokens and $25 per million output tokens, while Claude Sonnet 4.6 costs $3 and $15, respectively, according to the company's website.That puts Composer 2 at roughly one-tenth the cost of Opus 4.6 and about one-sixth the cost of Sonnet 4.6 on both input and output tokens. Users on X have added to the debate, with some praising the performance of Kimi after learning that Composer 2 was built on top of it."As someone who basically lives in opus 4.6, seeing an open-weight kimi 2.5 fine-tune actually beat it on coding benchmarks is wild," one X user wrote in response to Fynn's post. "Well that's a sign for RL Chinese is in new game," another user wrote, referring to reinforcement learning.Others were more critical of Cursor's handling of the disclosure, questioning why the company did not acknowledge Kimi upfront. "Cursor is becoming a model routing layer, not an IDE. they pick the cheapest model that clears a quality bar per task, wrap it in their UX, and pocket the margin," one user who goes by aira wrote on X.