Cursor, a coding platform developed by the startup Anysphere, has launched Composer, its inaugural in-house large language model (LLM). This release is part of the significant Cursor 2.0 platform update and aims to enhance the efficiency of coding tasks in production environments.
Composer has already been adopted by Cursor”s engineering team for daily development tasks, reflecting its advanced capabilities. The model reportedly completes most tasks in under 30 seconds while demonstrating robust reasoning skills across extensive and intricate codebases. According to the company, Composer operates at a speed four times greater than that of comparable intelligent systems and is specifically designed for “agentic” workflows, where autonomous coding agents can collaboratively plan, write, test, and review code.
Previously, Cursor”s “vibe coding” feature utilized leading proprietary LLMs from companies like OpenAI, Anthropic, and Google to assist users, including those without programming expertise, in writing or completing code based on natural language instructions. These options remain available for users.
Benchmarking and Performance
Composer”s performance is evaluated through an internal suite known as “Cursor Bench,” which is derived from real developer requests. This benchmark assesses not only the correctness of coding outputs but also the model”s compliance with established coding standards and engineering practices. Composer has achieved frontier-level coding intelligence, generating at a rate of 250 tokens per second, which is approximately twice the speed of leading fast-inference models and four times faster than similar frontier systems.
Cursor categorizes its models into several groups for comparison: “Best Open,” “Fast Frontier,” “Frontier 7/2025,” and “Best Frontier.” Composer aligns with mid-frontier systems in intelligence while delivering the highest recorded generation speed across all tested categories.
Innovative Development Approach
Research scientist Sasha Rush from Cursor shared insights into the development of Composer, emphasizing its architecture as a reinforcement-learned mixture-of-experts (MoE) model. Rush stated, “We used RL to train a big MoE model to be really good at real-world coding, and also very fast.” The co-design of Composer and the Cursor environment was crucial for ensuring efficient operation at production scale.
Unlike many machine learning systems, which can often be abstracted away from their full-scale applications, the team designed Composer to operate effectively in real-world coding environments. This approach involved training the model on actual software engineering tasks rather than relying on static datasets. Throughout the training process, Composer tackled specific challenges, such as producing code edits and generating explanations, using a variety of production tools.
The reinforcement learning loop was instrumental in optimizing both the accuracy and efficiency of the model. Composer has learned to make informed tool choices, utilize parallel processing, and avoid speculative responses. Over time, it has developed capabilities such as running unit tests and autonomously fixing errors.
Integration and User Experience
Composer is fully integrated into the Cursor 2.0 platform, which features a multi-agent interface that allows up to eight agents to operate simultaneously in separate workspaces. Within this system, Composer can function as one or more agents, either independently or in collaboration with others. Developers can review multiple outputs from these concurrent agent operations to select the most effective results.
The updates in Cursor 2.0 also include several enhancements to support Composer”s functionality. These features include an in-editor browser for testing code directly within the IDE, improved code review capabilities that compile changes across multiple files for efficient inspection, and sandboxed terminals for secure execution. Additionally, a voice mode has been introduced, allowing users to manage agent sessions via speech-to-text controls.
To train Composer effectively, Cursor developed a custom reinforcement learning infrastructure that combines PyTorch and Ray for asynchronous training across a significant number of NVIDIA GPUs. This infrastructure enables large-scale updates with minimal communication overhead, allowing for efficient training at low precision without the need for post-training quantization.
Enterprise users benefit from enhanced administrative controls over Composer and other agents, including team rules and audit logs. Cursor also offers various pricing tiers for individual users and businesses, accommodating different usage needs.
With Composer, Cursor is setting a new standard in AI-assisted programming, moving beyond mere suggestions to facilitate continuous collaboration between human developers and autonomous models. This innovative approach aims to redefine everyday programming by integrating autonomous systems directly into the coding workflow.
