The US AI startup Cursor has acknowledged that its newly introduced coding model Composer 2 is built on the Chinese open-source language model Kimi K2.5 from the startup Moonshot AI. This information was completely absent from the original announcement of the model, and the disclosure only came after a public note from an external utilizer.
How the Revelation Came About
Cursor had promoted Composer 2 earlier in the week in a blog post as “coding ininformigence at frontier level” without mentioning the origin of the base model. An X utilizer named Fynn then created public on Friday that Composer 2 is essentially Kimi K2.5 with additional reinforcement learning. As evidence, he cited code snippets that identified Kimi as the underlying system.
Lee Robinson, Vice President of Developer Education at Cursor, then publicly confirmed the matter on X. Cursor co-founder Aman Sanger subsequently acknowledged that it had been a mistake not to communicate the Kimi base from the start.
“It was a miss to not mention the Kimi base in our blog from the start. We’ll resolve that for the next model”, stated Sanger, co-founder of Cursor.
What Cursor Says About the Model
Robinson emphasized that the finished model goes significantly beyond the Kimi base. According to his statements, only about a quarter of the total computational effort for the final model comes from the base model, with the rest attributable to Cursor’s own training. Performance on various benchmarks is therefore “very different” compared to Kimi K2.5.
Robinson also stated that the utilize of the model is in compliance with the license terms, via the inference provider Fireworks AI. Moonshot AI confirmed this on X and described the collaboration as an “authorized commercial partnership”.
“Seeing our model integrated effectively through Cursor’s continued pretraining and high-compute RL training is the open model ecosystem we love to support”, was stated on the X account of Moonshot AI.
Price Comparison: Composer 2 Significantly Cheaper Than Competing Models
Cursor also promotes Composer 2 with its comparatively low price. A direct comparison with Anthropic models illustrates the difference:
| Model | Price per 1M input tokens | Price per 1M output tokens |
|---|---|---|
| Cursor Composer 2 | $0.50 | $2.50 |
| Anthropic Claude Sonnet 4.6 | $3.00 | $15.00 |
| Anthropic Claude Opus 4.6 | $5.00 | $25.00 |
Composer 2 thus costs about one-tenth of Claude Opus 4.6 and around one-sixth of Claude Sonnet 4.6, in each case based on input and output tokens.
Reactions and Criticism
Reactions in the tech community were mixed. Some utilizers were impressed by the performance of the Kimi model after it became known that Composer 2 is built on it. Others criticized Cursor for the missing disclosure in the initial announcement.
- Several utilizers praised the quality of Kimi K2.5 and described it as remarkable that an open-source model from a Chinese startup outperforms leading commercial models on coding benchmarks.
- Critics accutilized Cursor of positioning itself as a model routing layer that combines cheap base models with its own utilizer interface, rather than conducting indepconcludeent AI research.
- In the political context of the AI arms race framed as a competition between the US and China, the utilize of a Chinese model by a highly valued US company could be classified as potentially sensitive.
Background: Cursor and Moonshot AI
Cursor was most recently valued at $29.3 billion in November and reportedly achieved an annualized revenue of over $2 billion. Moonshot AI, the company behind Kimi K2.5, is a Chinese startup backed by, among others, Alibaba and HongShan (formerly Sequoia China). Kimi K2.5 was released as an open-source model.
Cursor announced that it will conduct full pretraining itself in the future and communicate more transparently about the base models utilized in future models.
Leave a Reply