Did Cursor really steal Kimi???

Cursor’s newly shipped “Composer 2” model is being treated as a major leap in coding performance-per-dollar—but a wave of scrutiny suggests it may not be as brand-new as its name implies. The controversy ignited after someone noticed “Kimmy K2.5” embedded in Composer 2’s API-related naming, prompting questions about whether Cursor effectively built its coding model by heavily post-training Moonshot AI’s openweight Kimmy base rather than starting from scratch.

The core claim emerging from the discussion is that Composer 2 is a refined, RL-tuned derivative built on top of Kimmy K2.5, with Cursor using large-scale compute and reinforcement learning to push coding ability toward “frontier” levels while keeping inference costs extremely low. In reported pricing terms, Composer 2’s default throughput is described as roughly 80–100 tokens per second at about $0.50 per million tokens in and $2.50 per million out—about 10x cheaper than Opus (quoted at $5 in and $25 out). External benchmarks cited in the conversation also place Composer 2 ahead of Opus on code-focused tests like TerminalBench 2, though the model still makes “dumb mistakes” that require steering.

What makes the situation more combustible is licensing and disclosure. Moonshot AI’s modified MIT license includes a commercial-use condition requiring prominent display of “Kimmy K2.5” in products or services exceeding certain scale thresholds (100 million monthly users or $2 million monthly revenue). Critics argue Cursor’s product naming and UI did not clearly surface that requirement. Speculation then turned to how Cursor might comply indirectly: Fireworks—described as a compute platform for hosting openweight models—provides hosted inference and RL tooling, and the discussion claims Cursor accessed Kimmy K2.5 through Fireworks as part of an authorized commercial partnership. In that framing, Fireworks’ pipeline allegedly handles the license-required disclosure “through inference partner terms,” allowing Cursor to avoid showing “Kimmy K2.5” directly in its own interface.

Supporters and defenders in the thread point to the technical reality that “openweight” models are commonly used as starting points. Post-training can drastically change behavior: the conversation emphasizes that many major model releases reuse the same pre-training base and then apply refinement to unlock new capabilities. The unusual part, according to this account, is the scale and outcome—using post-training on someone else’s model to reach coding results comparable to top-tier systems.

Beyond the immediate drama, the discussion frames a broader industry risk: if licensing ambiguity and perceived “rebranding” backlash deter openweight releases, smaller labs may hesitate to publish models at all. The long-term worry is that openweight cost improvements—credited here with driving better and cheaper frontier performance—could slow if model providers feel exposed to disputes over attribution, compliance, and “spirit” of licenses. The bottom line: Composer 2 appears genuinely strong and unusually economical for coding, but the path to it—starting from Kimmy K2.5 and the way that relationship was disclosed—has become a flashpoint for how openweight ecosystems may evolve under intense compute-price pressure.