Cursor’s new coding model, Composer 2, is essentially a heavily post-trained version of the open-weight Kimmy K2.5 model from Moonshot AI, but Cursor failed to clearly disclose this, raising ethical and licensing concerns. While Composer 2 offers impressive performance and cost-efficiency, the controversy highlights broader issues around transparency, attribution, and the sustainability of open-weight AI models in commercial use.
The video discusses the recent revelation that Cursor’s new coding model, Composer 2, is not entirely original but is actually built upon Kimmy K2.5, an open-weight model from Moonshot AI. While Composer 2 performs impressively, surpassing competitors like Opus in speed and cost-efficiency, the discovery sparked controversy because Cursor did not clearly disclose that their model was a heavily post-trained version of Kimmy K2.5. This lack of transparency raised questions about licensing compliance, especially given Kimmy’s modified MIT license, which requires prominent attribution for commercial products with large user bases or revenues—a condition Cursor seemingly skirted by using an inference partner, Fireworks AI, to obscure the model’s origins.
The video provides background on Cursor’s journey, highlighting its acquisition of Super Maven and the leadership of Jacob, who has focused on building AI models specialized in code autocomplete and generation. Composer 2 represents Cursor’s ambitious effort to create a bespoke coding model optimized for speed and cost, leveraging extensive reinforcement learning on top of Kimmy K2.5. This approach allowed Cursor to produce a model that is not only highly capable but also significantly cheaper to run—about ten times less expensive than competitors like Opus—making it attractive for enterprise use despite the challenges of maintaining profitability amid subsidized pricing from major AI labs like Anthropic.
A key point in the discussion is the economics of AI model inference, where companies like Anthropic heavily subsidize API usage, making it difficult for smaller players to compete on cost. Cursor’s strategy to use an open-weight model like Kimmy K2.5 as a base, then apply extensive post-training with their own data and compute resources, offers a way to reduce costs while maintaining high performance. However, this raises ethical and legal questions about the use of open-weight models in commercial products, especially when the original creators are not adequately credited or compensated, potentially discouraging smaller labs from releasing open-weight models in the future.
The video also delves into the technical aspects of model training, explaining that most major AI models undergo a two-step process: pre-training to build foundational knowledge and post-training (including reinforcement learning) to refine behavior and capabilities. Cursor’s Composer 2 is notable for achieving frontier-level performance primarily through post-training on Kimmy K2.5, a novel approach that blurs the lines between original and derivative work. This has sparked debate about the spirit of open-weight licensing and the responsibilities companies have when building commercial products on top of community-developed models.
In conclusion, while Composer 2 is a powerful and cost-effective coding model, the controversy around its origins highlights broader issues in the AI ecosystem regarding transparency, licensing, and the sustainability of open-weight models. The video’s creator expresses disappointment in how Cursor handled the disclosure but acknowledges the complexity of the situation and hopes for better practices in the future. The incident serves as a cautionary tale for the industry, emphasizing the need for clearer licensing terms and more ethical collaboration to ensure that innovation continues without undermining the contributions of smaller AI labs.