Cursor's Composer 2 Model Built on Kimmy K2.5 Sparks Disclosure Debate and Licensing Scrutiny

Cursor recently launched Composer 2, a new AI model designed specifically for coding tasks, garnering attention for its impressive performance metrics and cost efficiency. Internal benchmarks and external evaluations, including Terminal Bench 2, suggest Composer 2 outperforms leading models like Anthropic’s Opus in certain code-centric tasks, delivering 80-100 tokens per second at an estimated cost ten times lower than Opus. This development follows Cursor’s long-standing efforts to develop proprietary models, driven by the escalating inference costs of frontier models from providers like Anthropic and OpenAI, which are currently engaged in aggressive subsidization strategies. Cursor, leveraging its extensive chat history data—amassed from user interactions in its application—has pursued reinforcement learning (RL) and post-training to optimize models for code generation.

However, the excitement around Composer 2 quickly turned to scrutiny when a developer discovered via API URLs that the model was based on “Kimmy K2.5 RL 0317 S515 fast,” an open-weight model from Moonshot AI. This revelation sparked a debate regarding proper disclosure and adherence to Kimmy K2.5’s modified MIT license, which mandates prominent display of the model’s name for commercial products exceeding specific user or revenue thresholds. Initially, Moonshot AI team members expressed confusion, having no prior knowledge of Cursor’s integration. Cursor, through its engineering lead, later clarified that Composer 2 began from an open-source base, utilizing Kimmy K2.5 as the foundation for significant post-training, involving an estimated three times the compute used for K2.5’s original training. Cursor asserted compliance with the license through its inference partner, Fireworks AI, which reportedly handles the necessary disclosures. While the technical achievement of significantly enhancing an open-weight model for specialized coding tasks is acknowledged, the lack of upfront transparency from Cursor has drawn criticism from the broader AI community, raising concerns about the ‘spirit’ of open-weight licensing and its potential impact on future open model releases from smaller labs.