Cursor is clearing the air after its latest coding model, Composer 2, sparked quiet curiosity—and then loud debate—among developers.
When the model first rolled out, the company didn’t say much about what was powering it under the hood. That didn’t go unnoticed. Developers began poking around, checking API traces and model names, and soon enough, clues started pointing toward Moonshot AI’s Kimi K2.5.
That’s when the conversation picked up pace. Some questioned why it wasn’t mentioned earlier, while others wondered if there were deeper issues around usage or licensing. For a short time, even Moonshot AI’s founders weighed in publicly, raising concerns before later taking those posts down.
Cursor CEO Aman Sanger eventually addressed the situation in a more straightforward way. He admitted the company should have been upfront about using Kimi K2.5 from the beginning. According to him, the team had tested several models, but this one stood out in terms of performance. He also explained that Composer 2 isn’t just a base model—it’s been further refined with additional training and reinforcement learning to push its capabilities further.
Behind the scenes, the system runs on Fireworks AI’s infrastructure, which helps handle both the heavy lifting of inference and the fine-tuning process.
Interestingly, Moonshot AI has since taken a more positive stance, congratulating Cursor and framing the launch as a good example of how open models can evolve when other teams build on top of them.
As for Composer 2 itself, Cursor is positioning it as its most advanced coding model yet. It supports a large context window—around 200,000 tokens—and is built for more complex, agent-style workflows. That includes working across multiple files, handling terminal commands, and using different tools in a more seamless way.
The company also points to strong benchmark results, saying the model performs well on tests like Terminal-Bench and SWE-Bench Multilingual. In some cases, it even edges past competing systems, while still keeping costs relatively in check—a combination that’s likely to appeal to developers looking for both performance and efficiency.
Also Read: Figma Expands AI Integrations to Link Coding Tools with Design Workflows
