Cursor shipped Composer 2 on March 19 as its most capable proprietary coding model, claiming it beat Claude Opus 4.6 in benchmarks, but a developer discovered within 24 hours that the model wasn't built from scratch. A leaked API model identifier revealed the truth: accounts/anysphere/models/kimi-k2p5-rl-0317-s515-fast. That's Kimi K2.5, an open-weight model from Moonshot AI, fine-tuned with reinforcement learning (a training technique that uses reward signals to improve model behavior). Cursor had not mentioned this foundation anywhere in its announcement. How Did a Developer Uncover the Hidden Model Identity? A developer named Fynn was testing Cursor's OpenAI-compatible API endpoint when the internal model identifier leaked through the system response. The model ID itself told the entire story. Breaking down the components reveals what Cursor was actually using: - anysphere: Cursor's parent company, indicating this is an internal Cursor variant - kimi-k2p5: Kimi K2.5, Moonshot AI's publicly available open-weight model released under a modified license - rl: Reinforcement learning, the post-training method Cursor applied to the base model - 0317: March 17, 2026, likely the date this version was trained - s515: An internal version or experiment identifier used by Cursor's engineering team - fast: The optimized serving variant designed for quick inference Cursor's announcement emphasized "continued pretraining" and "scaled reinforcement learning" as innovations driving Composer 2's performance. The company's blog post, release notes, and model documentation made no mention of Kimi K2.5 as the foundation. What Does Cursor's License Violation Actually Mean? Moonshot AI released Kimi K2.5 under a Modified MIT License with a critical commercial clause. The license states that if a company uses the model in products generating more than $20 million in monthly revenue or serving more than 100 million monthly active users, it must "prominently display 'Kimi K2.5' on the user interface of such product or service". Cursor's numbers trigger this requirement multiple times over. The company reported roughly $2 billion in annualized revenue as of February 2026, translating to approximately $167 million monthly, which is 8 times the $20 million threshold. Cursor's user base includes over 1 million daily active users, with monthly active users likely well exceeding the 100 million requirement. Yet Cursor's interface displays "Composer 2" with zero mention of Kimi. Two Moonshot employees initially confirmed on social media that Cursor wasn't licensed for this use, then deleted their posts. Yulun Du, Moonshot AI's head of pretraining, publicly confirmed the tokenizer similarity and questioned Cursor's compliance with licensing terms. How Did Cursor Respond to the Accusation? Within 24 hours of the leak, Cursor acknowledged the foundation. The company released a statement confirming that Composer 2 started from Kimi K2.5 as an open-source base. Cursor claimed that only approximately 25 percent of the compute spent on the final model came from the base, with the remaining 75 percent derived from their own reinforcement learning training. This represents a genuine technical contribution on top of the open-weight foundation. Cursor also acknowledged the attribution gap, stating: "Agree with the feedback we should have mentioned the base up front, we will do that for the next model." The company indicated it is complying with the license through its inference partner, Fireworks AI. Moonshot AI confirmed the license is being followed, and the dispute appears resolved. However, the resolution came only after @fynnso leaked the model ID. Without that discovery, users would still not know what model they were paying for. Cursor's promise to disclose the base model upfront in future releases is welcome, but it addresses the future, not why this launch was different. What Does This Mean for Open-Source AI Licensing? The Cursor situation reveals both the strength and fragility of open-source licensing in the AI era. Moonshot AI wrote a Modified MIT License specifically anticipating this scenario: a well-funded company taking an open-weight model, fine-tuning it, and selling it without attribution. When Cursor triggered the license's commercial thresholds, community pressure forced transparency, and the company complied. The system worked, but not automatically. It required a developer to probe an API endpoint, discover the hidden model ID, and publicly call out the discrepancy. Open-weight licenses are enforceable, but transparency still has to be demanded rather than volunteered. Cursor's Composer 2 pricing undercuts industry standards significantly. The model costs $0.50 per million input tokens, massively undercutting standard frontier model pricing. The company claims Composer 2 scored 61.3 percent on CursorBench, beating Claude Opus 4.6, though CursorBench is a proprietary evaluation heavily adjusted to favor models running inside Cursor's environment, making direct comparisons to general-purpose models inherently flawed. The broader lesson is clear: open-weight models are reshaping AI development economics, but attribution and licensing compliance require vigilance. A $29 billion company built a highly capable, incredibly cheap coding agent by optimizing a model for its specific application. The community debate proves that self-reported benchmark scores matter much less than how well a tool actually functions in a developer's daily workflow. But it also proves that without transparency, users cannot make informed choices about what they are actually paying for.
