Cursor unveils budget-friendly coding model with Chinese origins

Cursor has now confirmed that its latest coding model has roots in China — a detail that was initially left out.

Over the weekend, Cursor executives shared in several X posts that Composer 2 was first developed using Kimi K2.5, an open-source model from Chinese startup Moonshot AI.

"We’ve assessed numerous foundational models based on perplexity metrics, and Kimi k2.5 stood out as the strongest!" stated Cursor's cofounder Aman Sanger on X on Saturday.

"Not mentioning the Kimi base model in our initial blog was an oversight," he commented further.

The revelation seems to have been triggered by an X user, Fynn, who mentioned on Friday that Composer 2 is essentially "just Kimi 2.5" supplemented by additional reinforcement learning.

To back up the assertion, the user referred to code snippets that seemed to indicate Kimi as the underlying framework.

'At least change the model ID name,’ the user remarked.

In reply to the user's post on X, Cursor's vice president of developer education, Lee Robinson, confirmed that Composer 2 was developed by building on the Kimi K2.5 open-source base.

"We plan to conduct full pretraining in the future," Robinson stated.

"About one-fourth of the computational resources used for the final model came from the base; the remainder results from our training," he added.

Robinson also noted that the company adheres to the model's licensing requirements through its service provider.

The Chinese startup stated on X on Saturday that Cursor is utilizing Kimi K2.5 under a legitimate commercial partnership.

"Seeing our model effectively integrated through Cursor's ongoing pretraining & intensive RL training matches the open model ecosystem we support," read their statement.

As of November, Cursor’s valuation was $29.3 billion.

In a blog post on Thursday, Cursor announced that Composer 2 is "advanced in its coding capabilities" and costs $0.50 per million input tokens and $2.50 per million output tokens, describing it as "a groundbreaking blend of intelligence and cost efficiency."

In 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 their website.

This makes Composer 2 roughly one-tenth the cost of Opus 4.6 and approximately one-sixth the cost of Sonnet 4.6 for both input and output tokens.

Discussions among X users continue, with some applauding the performance of Kimi after realizing Composer 2 was built on its framework.

"As someone who regularly uses opus 4.6, seeing an open-weight kimi 2.5 fine-tune outperform it on coding benchmarks is astonishing," commented one X user in response to Fynn’s post.

"That certainly indicates reinforcement learning in China is on a whole new level," another user noted.

Some, however, have critiqued Cursor's approach to the disclosure, questioning why the company did not initially credit Kimi.

"Cursor is evolving into a model routing platform, not just an IDE. They choose the most cost-effective model that meets quality standards for each task, integrate it into their UI, and profit from the difference," observed a user known as aira on X.