Cursor’s Composer AI codes 4x faster than competing models

Cursor has released Composer, a new AI agent model designed for software engineering that achieves frontier-level coding performance while generating responses four times faster than similar models. The model was trained using reinforcement learning on real-world software engineering challenges in large codebases, optimizing it for interactive development workflows that keep programmers in their coding flow.

How it works: Composer is built as a mixture-of-experts (MoE) language model that specializes in software engineering through reinforcement learning in diverse development environments.

• During training, the model receives problem descriptions and must produce optimal responses—whether code edits, plans, or informative answers—using tools like file reading/editing, terminal commands, and codebase-wide semantic search.
• The reinforcement learning process actively trains the model for efficient tool use and parallel processing to maximize response speed for interactive development.
• The model autonomously learned useful behaviors during training, including performing complex searches, fixing linter errors, and writing and executing unit tests.

In plain English: Think of Composer as an AI coding assistant that learned by practicing on thousands of real programming projects. Just like a human developer, it can read files, edit code, search through large projects, and run commands—but it does all of this much faster because it was specifically trained to work efficiently with these tools.

The benchmark: Cursor created Cursor Bench, an evaluation system that measures real-world usefulness to software developers rather than just correctness.

• The benchmark consists of actual agent requests from Cursor’s engineers and researchers, paired with hand-curated optimal solutions.
• It evaluates not only correctness but also adherence to existing codebase abstractions and software engineering best practices.

Technical infrastructure: Building Composer required significant investment in custom training infrastructure and systems research.

• The team built custom training infrastructure using PyTorch and Ray to enable asynchronous reinforcement learning at scale.
• They developed MXFP8 MoE kernels combined with expert parallelism and hybrid sharded data parallelism, allowing training across thousands of NVIDIA GPUs with minimal communication costs.
• Training with MXFP8 enables faster inference speeds without requiring post-training quantization.

In plain English: Cursor had to build specialized computer systems to train Composer, similar to how a car manufacturer needs custom assembly lines to build vehicles. They created new software tools that let thousands of powerful computer processors work together efficiently, which allows the AI to respond quickly without needing additional optimization after training.

Training environment: The reinforcement learning process required running hundreds of thousands of concurrent sandboxed coding environments in the cloud.

• Cursor adapted existing infrastructure from their Background Agents project, rewriting their virtual machine scheduler to handle the bursty nature and scale of training runs.
• This enabled seamless integration between RL training environments and production environments.

Performance results: Composer demonstrates competitive performance across different model categories in Cursor’s internal benchmarks.

• The model matches “Fast Frontier” models (like Haiku 4.5 and Gemini Flash 2.5) designed for efficient inference.
• It outperforms recent open-weight models such as Qwen Coder and GLM 4.6 (“Best Open” category).
• While it doesn’t surpass the absolute best frontier models like GPT-5 and Sonnet 4.5, it delivers comparable results at significantly higher speeds.

Real-world adoption: Cursor employees have been using Composer for their daily software development work in recent weeks, validating its practical utility for the workflows it was designed to optimize.