Cursor's Composer 2 model revealed to be Kimi K2 with reinforcement learning — a comms failure more than a strategy failure
Mar 20, 2026
Key Points
- Cursor's Composer 2, presented as a breakthrough model, was built by fine-tuning open-source Kimi K2 with reinforcement learning—a legitimate technique Cursor didn't disclose upfront.
- Lee Robinson claims three-quarters of Composer 2's compute came from Cursor's own training work, with licensing obligations met through inference partner agreements.
- Cursor's ARR continues growing despite the model ID leak and brand damage, suggesting the market is expanding fast enough to absorb the credibility hit.
Summary
Cursor announced Composer 2 yesterday as a breakthrough model with strong internal benchmark performance. A developer quickly identified the model ID in Cursor's OpenAI base URL, revealing it as Kimi K2 with reinforcement learning applied, not a from-scratch build.
Cursor's cofounders and Lee Robinson, who leads developer relations, confirmed the finding. Robinson said Composer 2 started from Kimi K2 as an open-source base, then applied reinforcement learning. He claimed only one quarter of the final model's compute came from the base model, with the remaining three-quarters from Cursor's own training work. Robinson stated they are "following the license through our inference partner terms."
The problem is framing, not strategy. Using an open-source model as a starting point and fine-tuning it with reinforcement learning is legitimate and widely practiced. Perplexity has done similar work, and Anthropic-derived models have been distilled and optimized by other labs. Cursor's mistake was positioning Composer 2 as a major architectural achievement without clearly disclosing the base model upfront. Robinson's language about "continual pre-training" obscured what was actually a fine-tuning effort.
The licensing question remains partially open. Some observers note that attribution obligations typically apply when building on open models, though Robinson claims compliance through their inference partner agreement. Others suspect full disclosure may already exist in Cursor's terms of service or Composer 2 documentation and simply was not surfaced in initial coverage.
On whether this matters to the business, the answer appears to be no. Cursor's ARR continues to grow despite the model ID leak and shifting market sentiment. The market is expanding fast enough that product momentum may absorb any brand damage. Jack Morris noted that Cursor is likely already planning Composer 3, which would be the first model fully pretrained from scratch for coding, suggesting the company views its current reliance on open-source bases as temporary.
Cursor did the smart technical thing but presented it poorly. It is a communications failure, not a strategy failure.