Cursor has introduced Composer, its first proprietary large language model (LLM) tailored for its integrated development environment (IDE). This innovative model is designed for agentic workflows, enabling autonomous AI agents to efficiently plan, write, and test code. Composer, a mixture-of-experts (MoE) model, enhances resource efficiency by activating only relevant experts for specific tasks. Although not open source, it’s fine-tuned using reinforcement learning (RL), leveraging tools like file editors and semantic search to solve real coding problems, which has led to emergent behaviors like complex code searches and automatic linter fixes.
With a processing speed of 250 tokens per second, Composer significantly surpasses comparable models in terms of responsiveness. Co-designed with its platform, it enhances development productivity by letting developers run multiple AI agents simultaneously. The assessment tool, Cursor Bench, measures not just code correctness but adherence to engineering practices, ensuring Composer provides frontier-level intelligence for software development. This positions it competitively in the growing LLM market.
Source link
