Open-source uproar: Cursor's Composer 2 accused of repackaging Kimi K2.5, licensing row erupts

Allegations and evidence

A fast-escalating controversy has engulfed AI startup Cursor after it launched Composer 2, a new programming model that community members say is largely built on the Chinese open-source model Kimi K2.5. It has been reported that a developer debugging Cursor’s API captured a model ID — "kimi-k2p5-rl-0317-s515-fast" — and that independent tests found Composer 2’s tokenizer identical to Kimi’s. Elon Musk reportedly commented “Yeah, it’s Kimi 2.5,” amplifying attention across the AI ecosystem overnight.

Admissions, license terms and money

Cursor initially described Composer 2 as “based on open-source work” and emphasized heavy in-house training, with one team member saying roughly one-quarter of final compute came from the base model and three-quarters from Cursor’s own RL training. It has been reported that Cursor later explicitly acknowledged KIMI K2.5 as the starting point. The dispute centers on licensing: Kimi’s license reportedly requires visible attribution if a derivative product’s monthly revenue exceeds $20 million — and Cursor, with annual revenue reportedly above $2 billion (monthly about $166 million), would be well above that threshold. How will investors and potential acquirers react? The question now is whether attribution alone will satisfy the community or whether deeper remedies are required.

Community reaction and broader implications

The open-source AI community and influencers piled in, calling the episode “profit arbitrage” on others’ R&D. Kimi (Kimi), whose K2.5 model is widely ranked among leading open models, reportedly responded graciously and framed the collaboration as a proud validation of its work, even as debate continues about compliance with the license terms. For Western startups that rely on Chinese open models, this raises fresh geopolitical and legal sensitivity: IP, licensing enforcement, and traceability of training data are drawing greater scrutiny amid trade-policy tensions and export controls that already complicate Sino-Western AI cooperation.

Technical context and stakes

K2.5 is reported to be a trillion-parameter Mixture-of-Experts model that activates roughly 32 billion parameters per inference, supports a 256K context window, and natively handles text, image and video — features that make it particularly well-suited to coding assistants that tackle large codebases and long chains of reasoning. With Cursor valued at roughly $29.3 billion and seeking much larger rounds, and Kimi’s backers having a reported valuation of about $4.3 billion, this is not just a reputational skirmish. It is a test of how open-source licensing, corporate incentives and global politics will interact as AI commoditizes foundational models. Who enforces the rules when the currency is model weights and the payoff is hundreds of millions in monthly revenue?