Cursor Launches Composer, a Fast-Track LLM for Coding Tasks

The coding platform Cursor, developed by startup Anysphere, has unveiled its first in-house large language model (LLM), named Composer, as part of the significant Cursor 2.0 update. This new tool promises to deliver a remarkable fourfold increase in speed for coding tasks, enhancing the efficiency of programming in production-scale environments.

Composer is currently in active use by Cursor’s engineering team, which highlights its reliability and maturity. According to the company, this model can complete most coding tasks in less than 30 seconds while maintaining a high level of reasoning across large and complex codebases. By optimizing for “agentic” workflows, Composer allows autonomous coding agents to collaboratively plan, write, test, and review code.

Previously, Cursor relied on established LLMs from leading organizations like **OpenAI**, **Anthropic**, **Google**, and **xAI** to support its “vibe coding” approach—enabling users to generate code using natural language instructions. While these options remain available, Composer’s launch marks a notable evolution in Cursor’s capabilities.

Benchmarking Composer’s Performance

Composer’s performance has been rigorously evaluated through an internal benchmarking suite called “Cursor Bench,” which assesses its abilities based on real developer requests. This benchmark evaluates not only the correctness of code but also the model’s adherence to established coding standards and practices.

In these tests, Composer demonstrates frontier-level coding intelligence, generating output at a rate of 250 tokens per second—about twice as fast as other leading models and four times faster than comparable systems. The results categorize Composer alongside models deemed “Best Frontier,” indicating its competitive edge in the market.

Research scientist **Sasha Rush** has shared insights into Composer’s development on the social network X, explaining that the model was created using reinforcement learning and a mixture-of-experts architecture. He stated, “We used RL to train a big MoE model to be really good at real-world coding, and also very fast.” This design allows Composer to operate efficiently in production settings, addressing complex coding challenges directly.

Training and Integration into Cursor 2.0

The development of Composer builds upon an earlier prototype known as Cheetah, which focused on low-latency inference for coding tasks. Rush noted that while Cheetah was primarily a speed test, Composer not only matches that speed but enhances reasoning and generalization capabilities.

With Composer’s integration into Cursor 2.0, the platform introduces a multi-agent interface, allowing up to eight agents to operate simultaneously in isolated workspaces. This architecture enables Composer to perform tasks independently or in collaboration with other agents, providing developers with options to compare multiple outputs from concurrent runs.

Cursor 2.0 also includes several features designed to enhance Composer’s functionality. Notable updates include an in-editor browser that facilitates code testing within the integrated development environment (IDE), improved code review processes, and secure execution of agent-run shell commands through sandboxed terminals. Additionally, a new voice mode allows users to manage agent sessions using speech-to-text controls.

To support the training of Composer at scale, Cursor developed a specialized reinforcement learning infrastructure using **PyTorch** and **Ray**. This system allows for asynchronous training across thousands of **NVIDIA** GPUs, optimizing performance and efficiency in real-time coding tasks.

As Composer is designed for real-world application, it is built to function within the same coding environment as its users, effectively managing version control and iterative testing. This alignment with actual coding conditions is crucial for achieving reliability in software development.

With its rollout, Composer is poised to transform the coding landscape by offering an AI tool that not only accelerates task completion but also enhances the quality of coding through intelligent integration. The pricing for individual users ranges from a free basic tier to a $200 per month ultra plan, while enterprise solutions begin at $40 per user monthly.

In summary, with Composer, Cursor is not just introducing a faster coding model; it is advancing the capabilities of AI in software development. By combining reinforcement learning with practical integration into existing workflows, Composer sets a new standard for coding assistance, promising a more collaborative future for developers and AI systems.