Cursor is building a frontier model from scratch, and the move redefines what a coding tool company is
At its inaugural Compile conference in June 2026, Cursor revealed a 1.5-trillion-parameter model trained from scratch on over 100,000 GPUs. The announcement marks a deliberate transformation: a product company stepping into the business of frontier AI research.
Cursor CEO Michael Truell stood at the company’s first Compile conference in San Francisco on June 16, 2026, and announced something that most coding tool companies never attempt: a model built entirely from scratch, at frontier scale, with no third-party base.
The model, named Composer 3 in several technical readouts of the event, carries more than 1.5 trillion parameters and was pre-trained on over 100,000 GPUs. According to reporting from AI Insiders and Tech Times, the training run was conducted on the Colossus supercomputer cluster in Memphis, though coverage of the event attributes that infrastructure variously to xAI and to SpaceX, and the record on that point remains unsettled. What is not in dispute is the scale: 10 to 20 times the compute of any prior Cursor model, according to multiple accounts of the Compile keynote.
The design choices signal a clear break from the company’s prior approach. As noted by AIScroll and Alpha Signal in their readouts of the announcement, the model is trained without the Kimi base that underpinned earlier versions, a departure that reflects a decision to own the full stack rather than build on an existing foundation. More telling still is the positioning: Truell described the model as generally intelligent, not a specialized coding system. That distinction matters. A general-purpose model at this scale, shipped through a coding interface, is a different product category from the one Cursor occupied before.
New cursor model being teased at compile. Same size as Claude Opus and GPT55. Trained from scratch. No more Kimmy Base. 10 to 20x more compute versus composer generally intelligent not just coding releases in the next couple of weeks. Nathaniel Whittemore
According to The AI Opportunities, the evolution is legible in the company’s own usage data. Three years ago Cursor was primarily an editor. By the time Truell took the Compile stage, 95 percent of users were engaging with it as an agent. The infrastructure investment is the company catching up to how people were already using the tool.
The timeline Truell set was aggressive. Multiple outlets reported that he expected the model to reach users within weeks of the June 16 announcement. Nathaniel Whittemore, founder and chief executive of Superintelligent, flagged the announcement a day after the conference, noting the size comparison to Claude Opus and GPT-5.5, the break from the prior base model, the compute multiple, and the general-intelligence framing. His summary, made on June 17, anticipated a release within the next couple of weeks. That window closed on July 1, 2026.
The call itself was close to the reported facts but not a precise match. Whittemore’s account of the size comparison and the compute ratio tracked the Compile announcement accurately. His “next couple of weeks” framing was consistent with Truell’s “within weeks” language, though neither phrasing maps to an exact date, and the gate required resolution by July 1. The fit is real, but it reflects the same loose timeline the announcement itself carried.
What the Compile conference makes plain is that Cursor has made a structural bet. Training a 1.5-trillion-parameter model from scratch is not an incremental product decision. It requires a level of capital commitment, infrastructure access, and research capacity that puts the company in direct competition, at the model layer, with the labs whose outputs it had previously relied on. Whether the model performs at the level its scale implies is a question the coming months will answer. What is already settled is the direction: Cursor is no longer a product built on top of frontier AI. It is attempting to be one.