Cursor's New AI Model Built on Chinese Open-Source Code

Cursor's New AI Model Built on Chinese Open-Source Code

A significant development emerged this week in the AI coding assistant space, accompanied by a notable omission in its initial announcement. Cursor, the high-profile startup, unveiled its latest model, Composer 2, heralding it as a frontier-level coding intelligence system.

However, the developer community quickly raised questions about the model's origins. A user on X, Fynn, presented evidence suggesting Composer 2 was not built from the ground up but was instead based on Kimi 2.5, an open-source model recently released by Chinese firm Moonshot AI. Code snippets appeared to reference Kimi, leading to public scrutiny and comments about the lack of model ID renaming.

This revelation carried extra weight given Cursor's substantial U.S. backing and valuation, contrasting with the base model's Chinese origin. The company's initial announcement made no mention of Moonshot AI or the Kimi model.

Cursor Acknowledges the Open-Source Foundation

Cursor's response came from Lee Robinson, Vice President of Developer Education. He confirmed that Composer 2 did indeed start from an open-source base. Robinson clarified the scale of additional work involved, stating only about a quarter of the total compute spent on the final model came from that base, with the remainder dedicated to their own extensive training. He emphasized that this subsequent training resulted in benchmark performance distinctly different from the original Kimi model.

Licensing and Partnership Details Clarified

On the critical question of licensing, Robinson asserted their use was fully compliant. This point was reinforced by the official Kimi account on X, which congratulated Cursor and specified the use occurred through an authorized commercial partnership with Fireworks AI. The Kimi team expressed pride in seeing their model serve as a foundation, framing it as a positive example of the open model ecosystem they support, where continued pretraining and reinforcement learning can build effectively upon existing work.

The Unspoken Geopolitical Context

The situation inevitably touches on the competitive landscape of global AI development. The decision to not initially credit the Chinese base model may stem from broader tensions, where the advancement of AI is often portrayed as a strategic race between nations. The success of models like DeepSeek from China has previously stirred competitive concern in Silicon Valley, making any perceived reliance a sensitive topic.

Cursor co-founder Aman Sanger addressed the oversight directly, calling it a miss to not mention the Kimi base in their initial blog post and committing to correct this in future announcements.

A Testament to the Evolving Model Ecosystem

This incident highlights a growing trend in AI development where building upon powerful open-source foundations is becoming standard practice. The focus shifts from pure originality to the value added through specialized training, fine-tuning, and alignment. The dialogue between Cursor and Moonshot AI underscores a collaborative, if complex, international ecosystem where innovation is increasingly iterative and interconnected.

Related articles