Cursor is training its own Opus-class model from scratch. They are no longer fine-tuned on open-source model like Kimi K2. This is a major step up in capability and independence. They are using the 10-20X compute provided by xAI and Colossus 2. Screenshot Screenshot In a few weeks the 1.5 trillion parameter coding model and harness. Then in a month or two 6T and 10T parameter models.Screenshot Composer 2.5 is already competitive with top models (Opus 4.7 / GPT-5.5 level) but is cheaper and more efficient. It was built on a China’s open-source Kimi K2 and improved with heavy reinforcement...