Cursor is CAUGHT red handed...
TL;DR
Cursor’s new Composer 2 appears to be built on Kimi K2.5 — the controversy started when users spotted internal naming that pointed to Moonshot AI’s open model, and even Elon Musk chimed in with “Yeah, it’s Kimi 2.5.”
The real issue wasn’t theft, it was attribution — Wes says the evidence suggests Cursor used Kimi through Fireworks AI under an authorized commercial setup, but failed to clearly credit Moonshot despite Kimi’s modified MIT-style license requiring prominent disclosure for large companies.
Cursor says only about 25% of Composer 2’s compute came from the base model — according to Lee Robinson, roughly three quarters of the final model’s compute came from Cursor’s own post-training and reinforcement learning, which helps explain why Composer 2’s evals look meaningfully different.
Cursor’s technical contribution is a long-horizon coding trick called self-summarization — the model pauses mid-task, compresses its working context into around 1,000 tokens, and keeps going, which lets it handle projects that would otherwise blow past the context window.
Their showcase example is wild: getting Doom running on MIPS in 170 turns — Wes highlights Cursor’s blog post claiming an early Composer 2 checkpoint solved a Terminal-Bench 2.0 challenge by compressing more than 100,000 tokens into summaries while iterating toward a working port.
Wes thinks Cursor stayed quiet partly because “built on a Chinese model” is terrible optics right now — with US-China AI tensions high and enterprise buyers sensitive to Chinese dependencies, he argues the omission was likely a PR dodge, not a simple attempt to slap a new label on someone else’s work.
The Breakdown
The leak that set everything on fire
Wes opens with the punchline: Cursor launched Composer 2, people loved it, then one user noticed it was still being referred to internally as “Kimi K2.5.” That turned a hype cycle into instant drama, especially once Finn posted that Composer 2 looked like Kimi K2.5 plus reinforcement learning and Elon Musk replied, “Yeah, it’s Kimi 2.5.”
Why this matters: Cursor is huge, and Kimi isn’t just any model
He pauses to frame the stakes. Cursor is one of the fastest-growing AI software companies, reportedly valued near $30 billion and doing over $2 billion in annualized revenue, while Kimi K2.5 is a strong open model from China’s Moonshot AI, especially good at agentic tasks. The catch is the license: smaller users can treat it like open source, but big companies are supposed to prominently disclose that they’re using it.
Cursor’s first response felt carefully worded
Lee Robinson from Cursor confirmed Composer 2 started from “an open-source base,” but at first still avoided saying “Kimi” out loud. He said only about one quarter of the compute in the final model came from the base model, with the rest coming from Cursor’s own training and RL, and argued they were complying with the license through their inference partner’s terms.
Then Kimi’s side chimed in — and briefly escalated things
A Moonshot/Kimi employee, Yulun Du, posted that they tested Composer 2 and found it used the same tokenizer as Kimi, saying they were shocked Cursor had neither respected the license nor paid fees. That post later disappeared, and was replaced by a much calmer official statement from Kimi.ai congratulating Cursor and saying Fireworks AI was the inference provider in an authorized commercial partnership.
So what actually happened?
Wes’s reconstruction is pretty blunt: Cursor presented Composer 2 in a way that made many people assume it was trained from scratch, then only got more explicit after internet sleuthing forced the issue. His conclusion is that the outrage is mostly about failing to give proper credit to the open model community, not about theft or some secret illegal use.
Why Cursor may have stayed vague in the first place
He gives two reasons. First, Cursor doesn’t want to look like “just a wrapper” on top of someone else’s model, especially after a massive valuation. Second, saying your flagship coding model is built on top of a Chinese base model is a political and enterprise-sales headache in the middle of the US-China AI race, so even a full disclosure might have triggered a different kind of backlash.
The actually interesting part: Cursor’s self-summarization research
Wes then shifts from scandal to substance, and this is where he clearly thinks people are missing the story. Cursor’s blog describes a system where the model pauses during long coding tasks, summarizes what matters into roughly 1,000 tokens, and continues from that compacted state; the summaries themselves get reinforcement learning feedback, so bad summaries “slowly go extinct” while useful ones survive.
Doom, distillation jokes, and the bigger open-source takeaway
The flashy example is a Terminal-Bench 2.0 challenge: make 1993 Doom run on a MIPS-based setup from source code, which an early Composer 2 checkpoint reportedly solved over 170 turns after compressing more than 100,000 tokens of context. Wes also riffs on how everyone is building on everyone else — even joking via a tweet that a kid asking endless questions is “an advanced distillation attack” — before landing on Clement Delangue’s point: open source is now a competition enabler, and the frontier increasingly belongs not just to whoever trains from scratch, but to whoever adapts, fine-tunes, and productizes fastest.