Cursor launched Composer 2 last week, billing it as offering "frontier-level coding intelligence." The base model underneath it is Kimi k2.5, an open source release from Moonshot AI, a Chinese company backed by Alibaba and HongShan (formerly Sequoia China).

Cursor VP Lee Robinson disclosed the arrangement after the launch. According to Robinson, roughly 1/4 of the compute behind Composer 2 came from the Kimi k2.5 base model. Cursor contributed the remaining 3/4 through continued pretraining and high-compute reinforcement learning training. The company accessed Kimi k2.5 through an authorized commercial partnership with Fireworks AI.

What Cursor Actually Built

The 1/4 base, 3/4 custom-training split matters for understanding what Composer 2 is. It is not a Kimi wrapper. Cursor put substantial compute into post-training, which typically drives most of the final behavior in coding-specific models. Whether that investment produced something materially better than fine-tuning a different base model is not addressed in available information.

Moonshot AI's Kimi k2.5 is open source, which makes this arrangement straightforward legally. The commercial partnership with Fireworks AI handled the compute access. Nothing unusual there.

The Business Context

Cursor raised $2.3 billion in fall 2025 at a $29.3 billion valuation. The company reportedly exceeds $2 billion in annualized revenue. Building on an open source base model rather than training from scratch makes financial sense at any valuation, but it is a notable choice for a product positioned as frontier-level.

This could mean Cursor concluded the Kimi k2.5 base gave better returns on their training investment than alternatives. One possibility is that open source models from Chinese labs are now competitive enough to serve as a foundation for top-tier commercial coding products. The actual benchmark comparisons are not part of what Cursor has disclosed.

The Disclosure Timing

Cursor announced Composer 2 without leading with the Kimi k2.5 foundation. Robinson's disclosure came in response to questions. That is not unusual for model launches, but it is worth noting given the "frontier-level" framing and the geopolitical sensitivity around Chinese AI infrastructure in U.S. developer tools.

The product exists. The training split is disclosed. What is not yet available is independent evaluation of Composer 2 against the models it claims to match.

Source: Techcrunch