Composer 2.5

Product Introduction

1. Definition: Composer 2.5 is a specialized, advanced AI coding agent developed by Cursor, built upon the open-source Moonshot Kimi K2.5 large language model (LLM) checkpoint. It is technically categorized as a long-horizon, agentic AI model fine-tuned for complex software development and problem-solving tasks.
2. Core Value Proposition: Composer 2.5 exists to provide developers and enterprises with a significantly more intelligent and reliable AI collaborator capable of sustained, multi-step work on complex coding projects. Its primary value is a substantial improvement in intelligence, behavior, and effort calibration over its predecessor, particularly for long-running agentic tasks.

Main Features

1. Targeted RL with Textual Feedback: This feature addresses the credit assignment problem in long Reinforcement Learning (RL) rollouts. Instead of a single, delayed reward signal, the system injects localized textual feedback (e.g., a hint like "Reminder: Available tools are...") at the exact point in a task where the model could improve. It then uses on-policy distillation to align the model's behavior with a "teacher" distribution generated from the hinted context, providing a precise training signal for specific behaviors like tool usage, coding style, or communication.
2. Large-Scale Synthetic Data Generation: Composer 2.5 was trained with 25x more synthetic tasks than Composer 2. To create progressively harder challenges, the system uses techniques like "feature deletion," where the agent must delete and then reimplement specific features in a codebase while maintaining functionality, with tests providing verifiable rewards. This method grounds training in realistic codebases and continuously pushes the model's problem-solving limits.
3. Advanced Distributed Training Infrastructure (Sharded Muon & Dual Mesh HSDP): The model leverages a sophisticated, high-performance training stack. Sharded Muon with distributed orthogonalization applies momentum updates and Newton-Schulz orthogonalization at a granular level (per attention head, per expert) while efficiently managing sharded parameters across GPUs, achieving step times as low as 0.2 seconds for a 1-trillion parameter model. Dual Mesh Hybrid Sharded Data Parallelism (HSDP) uses separate parallelism layouts for non-expert and expert weights, allowing optimal communication overlap and compute distribution, avoiding wide communication for small parameters while efficiently scaling the massive expert components.

Problems Solved

1. Pain Point: The inability of previous AI coding agents to maintain coherent, correct behavior over long, complex tasks involving hundreds of steps or tool calls, leading to confusion, errors, and unreliable outcomes.
2. Target Audience: The primary users are professional software engineers, engineering teams, and tech leads working on large, complex codebases. Secondary users include researchers and developers requiring AI assistance for intricate, multi-step problem-solving beyond simple code generation.
3. Use Cases: Essential scenarios include refactoring large legacy code modules, implementing complex new features from detailed specifications, debugging intricate, multi-system issues, performing comprehensive code reviews across entire modules, and autonomously working on long-horizon projects like building a small application from scratch.

Unique Advantages

1. Differentiation: Compared to general-purpose LLMs (like GPT-4) or earlier coding agents, Composer 2.5 excels specifically in sustained, agentic work. It demonstrates superior "effort calibration" and follows complex instructions more reliably over long contexts, making it less prone to giving up or deviating from the task. Its pricing for the "fast" tier is also positioned as more cost-effective than comparable frontier model tiers.
2. Key Innovation: The integration of Targeted RL with Textual Feedback is a key innovation. It moves beyond monolithic reward models to provide surgical, context-aware training signals. This allows for the fine-tuning of specific, localized behaviors (e.g., avoiding a specific type of tool error) without retraining the entire reward model, leading to more nuanced and controllable model improvements.

Frequently Asked Questions (FAQ)

1. What is Composer 2.5 and how is it different from Composer 2? Composer 2.5 is Cursor's next-generation AI coding agent, offering a substantial improvement in intelligence and behavior over Composer 2, especially on long-horizon tasks. Key differences include training with targeted RL feedback, 25x more synthetic data, and advanced distributed training techniques, resulting in better instruction following and collaboration.
2. How much does Composer 2.5 cost to use? Composer 2.5 is priced at $0.50 per million input tokens and $2.50 per million output tokens. A faster variant with identical intelligence is available at $3.00/M input and $15.00/M output tokens. Cursor often provides promotional double usage credits for new releases.
3. What kind of tasks is Composer 2.5 best suited for? Composer 2.5 is specifically optimized for long-horizon, agentic coding tasks. This includes multi-file refactoring, implementing complex features from scratch, extensive debugging sessions, and autonomously working through lengthy project specifications where maintaining context and consistent behavior is critical.
4. What model is Composer 2.5 based on? Composer 2.5 is built on the same open-source foundation as Composer 2: Moonshot's Kimi K2.5 model checkpoint. Cursor then applies its proprietary fine-tuning, reinforcement learning, and synthetic data pipeline to create the specialized Composer agent.
5. What is "targeted RL with textual feedback" in Composer 2.5? It is a training technique that inserts short, corrective text hints (e.g., "Available tools are X, Y, Z") directly into the model's context at the point of a mistake during training. The model then learns from this localized feedback via distillation, allowing for precise correction of specific behaviors like incorrect tool calls or style violations, which are hard to correct with only a final task reward.