The rise of Cursor has been one of the most watched trajectories in the artificial intelligence sector, marking a shift in how developers interact with code. With a valuation reaching $2.3 billion and annualized revenues reportedly exceeding $2 billion, the company has transitioned from a niche startup to a heavyweight in the “frontier-level” coding intelligence space. However, recent discussions surrounding the launch of Composer 2 have sparked a broader conversation about the transparency of the open-source ecosystem and the strategic decisions behind building on international foundations.
Cursor recently achieved a multi-billion dollar valuation and significant annualized revenue. How does the decision to build on an open-source base like Kimi 2.5 rather than developing a proprietary foundation impact your long-term competitive moat and your capital allocation strategy for R&D?
Our strategy is built on the belief that speed and intelligence are more valuable to our users than the pride of building every single layer from scratch. By starting with an open-source base like Kimi 2.5, we can bypass the initial, massive capital drain of foundational pre-training and instead focus our R&D budget on specialized coding intelligence. This approach allowed us to reach a $29.3 billion valuation by demonstrating that our “moat” isn’t just the model itself, but our ability to refine and integrate it into a world-class developer workflow. We allocate our resources where they provide the most leverage, ensuring our $2 billion in annualized revenue is fueled by product performance rather than redundant infrastructure costs.
Only a portion of the compute for Composer 2 was dedicated to the base model, with the majority spent on additional training and reinforcement learning. Can you detail the specific technical steps involved in this refinement process and how you measure the resulting performance gap between the base and the final product?
The technical reality is that while the base model provides the linguistic foundation, only about one-fourth of the total compute for the final Composer 2 model came from that original base. The remaining three-quarters of our compute was dedicated to intensive continued pre-training and high-compute Reinforcement Learning (RL) specifically designed for complex software engineering tasks. We measure the gap through rigorous internal benchmarks that simulate real-world debugging and multi-file refactoring, which often show “very different” results compared to the raw base model. This refinement process transforms a general-purpose model into a specialized engine that understands the nuances of a developer’s specific intent and codebase.
The use of a Chinese-developed model as a foundation can be perceived as controversial in the current geopolitical “AI arms race.” What specific criteria do you use to evaluate the licensing and security of international open-source partnerships, and how do you navigate the transparency trade-offs when announcing such integrations?
We evaluate every partnership through a lens of technical excellence and strict legal compliance, ensuring that any model we use is part of an authorized commercial agreement. In the case of Kimi 2.5, we worked through an existing ecosystem and a partnership with Fireworks AI to ensure the integration was fully licensed and secure for our global user base. We recognize that the framing of an “arms race” creates sensitivities, and in hindsight, our co-founder Aman Sanger admitted it was a miss not to mention the Kimi base in our initial announcement. Our goal moving forward is to be more proactive in highlighting these collaborations, as we believe a global open-model ecosystem ultimately benefits the entire developer community.
While performance on benchmarks may differ significantly after post-training, the underlying model ID remained visible to some users. What does your internal workflow look like for rebranding and integrating third-party models, and what steps are being implemented to ensure more comprehensive disclosure in future product launches?
When users spotted the Kimi model ID within the code, it served as a clear signal that our internal workflow for model rebranding needed to be more robust and transparent. Our current process involves deep integration where we wrap the third-party foundation in our proprietary RL layers, but we clearly need to align our technical identifiers with our public-facing communications. We are refining our launch protocols to ensure that the “frontier-level” claims we make are backed by a clear disclosure of the model’s lineage from the very first blog post. We are committed to fixing this for the next model launch to ensure that our users feel fully informed about the technology powering their IDE.
Your partnership with Fireworks AI facilitated the commercial use of the Kimi foundation. How does leveraging an existing ecosystem of open models accelerate your deployment cycles, and what are the specific metrics you use to determine when a base model is “frontier-level” enough for your users?
Leveraging the open model ecosystem through partners like Fireworks AI allows us to cut months, if not years, off our deployment cycles by starting at a high baseline of intelligence. We determine if a base model is “frontier-level” by testing its raw reasoning capabilities and its ability to absorb our specific reinforcement learning without “forgetting” general logic. If a model can effectively support our high-compute RL training and result in a measurable leap in coding accuracy, we consider it a viable foundation. This collaborative approach is exactly the kind of ecosystem we love to support because it turns shared progress into specialized tools for our customers.
What is your forecast for the future of model development in the coding space?
The future of coding AI will move away from the “one model to rule them all” philosophy and toward a more modular, iterative approach where specialized training is the primary differentiator. We will see more U.S. startups utilizing global open-source foundations as a starting point, then investing heavily in proprietary reinforcement learning to create highly specialized agents. The winners in this space won’t necessarily be the ones with the largest raw compute for base training, but those who can most effectively refine existing models to handle the messy, complex reality of professional software engineering. Expect a shift toward total transparency in model lineage as the industry matures and recognizes that the value is in the refinement, not just the foundation.
