Cursor 3 Launches with a Custom Coding Model Trained from Scratch
Cursor 3, the latest version of the Anysphere IDE, has launched with a new custom coding model called Composer 2. This model was trained from scratch on a curated code corpus and optimized for specific IDE tasks, unlike fine-tuned or distilled models from large language models.
Why it matters
Anysphere's custom Composer 2 model and the new Cursor 3 features it enables represent a shift in the IDE market, where vendors are building their own specialized AI models to differentiate their products.
Key Points
- 1Composer 2 is Anysphere's own frontier coding model, trained from scratch for IDE tasks
- 2Cursor 3 includes several new features like parallel agents, design mode, and BugBot that leverage the custom model
- 3The custom model achieves 200+ tokens per second, much faster than APIs for large language models like GPT-5 and Claude
- 4Anysphere is the third IDE vendor to develop its own custom coding model, after Microsoft and Google
Details
Cursor 3 is not a single feature update, but a major release that includes several new capabilities built on top of Anysphere's custom Composer 2 coding model. Composer 2 was trained from scratch on a curated code corpus, with the training objective focused on specific IDE tasks like reading repositories, planning edits, running tests, and fixing issues. This is different from fine-tuning or distilling large language models like GPT-5 or Claude. The custom model achieves 200+ tokens per second on Anysphere's own GPU kernels, much faster than the 70-90 tokens per second typical of API access to these large models. This speed boost enables new Cursor 3 features like parallel agent runs, a Figma-style design mode, and an always-on BugBot that can automatically open pull requests to fix regressions. Anysphere is not the first IDE vendor to develop its own custom coding model - Microsoft and Google have done so as well, leveraging their ownership of the underlying infrastructure.
No comments yet
Be the first to comment