Cursor Unveils 1.5‑Trillion‑Parameter Model Trained on 100K GPUs After Musk’s Acquisition

Following SpaceX’s $60 billion purchase of Cursor, the startup revealed a new large‑scale model at its first flagship conference, Cursor Compile. The model contains roughly 1.5 trillion parameters and was pre‑trained on more than 100,000 GPUs.

CEO Michael Truell stated that the model’s size matches that of Opus 4.8 and GPT‑5.5, both of which remain under the 2 trillion‑parameter threshold. He suggested that, given the current performance of GPT‑5.5 and Opus 4.8, open‑source projects could eventually reach comparable capabilities.

Truell emphasized that scale is the only remaining moat in the AI race. To date, Anthropic is the sole lab that has successfully moved to the ~10 trillion‑parameter regime. He expressed doubt that OpenAI could close the gap by year‑end, noting that Anthropic’s continued investment in RL‑based training for its Mythos series should keep it ahead.

He also evaluated other leading labs: Google’s models have not been scaled to the same level and have pursued overly aggressive sparsity strategies, with limited translation of post‑training and RL improvements into stable products; OpenAI is still digesting the fallout from the GPT‑4.5 roadmap; and both xAI and Meta are preparing for the next wave of large‑model competition.

Some observers questioned the provenance of the “Opus‑scale and GPT‑5.5‑scale” claim, asking for evidence of the underlying parameter counts.

Beyond the headline numbers, Cursor’s strategic shift is evident: the company is moving from merely invoking external models to training its own foundational model. Three concrete changes differentiate the new model from previous versions: (1) it is larger, approaching the frontier of commercial LLMs; (2) it is trained from scratch rather than fine‑tuned from an open‑source base; and (3) the compute budget has been increased ten‑ to twenty‑fold compared with prior efforts, expanding from a modest Composer‑1/2.5 setup to a massive GPU cluster.

The team stresses that the model is not limited to code generation. Their vision is a more general‑purpose AI that can act as an engineering colleague—planning, testing software, interacting with UI elements, and clearly explaining any changes it makes. In other words, Cursor aims to evolve from pure code synthesis to full‑stack software‑engineering collaboration.

Truell indicated that training is already underway and that a public release is expected within the next few weeks, with external compute and infrastructure support likely supplied by SpaceX.

Ultimately, the true capabilities of the model will remain unknown until the system is released and evaluated.