Cursor Composer 2.5: AI Coding with 25x More Synthetic Training

Cursor has released Composer 2.5, its latest AI coding model optimized for long-horizon software engineering tasks. The model is built on Moonshot AI's Kimi K2.5 base and features significant improvements in agentic coding workflows.

Key Improvements

Training Scale

• 25x more synthetic tasks compared to Composer 2
• Trained with dynamic task selection that gets progressively harder during reinforcement learning
• Real-time RL pipeline running hundreds of thousands of sandboxed coding environments

Performance

• Handles long-running coding jobs that require hundreds of individual actions
• Maintains coherence across extended sessions where previous models lost context
• 60%+ cost reduction compared to frontier models at similar capability levels

Pricing

• Standard: $0.50/M input / $2.50/M output tokens
• Fast tier: $3.00/M input / $15.00/M output tokens
• Lower cost than comparable frontier models

Architecture Details

Composer 2.5 is a trillion-parameter mixture-of-experts (MoE) model based on Kimi K2.5:

• 32 billion active parameters
• Continued pretraining on code-dominated data
• Large-scale reinforcement learning with real Cursor sessions
• Custom low-precision kernels for efficient Blackwell GPU training

Use Cases

Ideal For:

• Multi-file refactoring projects
• Long-horizon feature development
• Complex debugging across codebases
• Automated test generation and maintenance

Example Workflow:

User: "Refactor this entire codebase to use async/await and add tests"
Composer 2.5: Plans, executes, and maintains context across hundreds of file changes

Training Methodology

1. Continued Pretraining: Code-focused data to deepen coding knowledge
2. Supervised Fine-tuning: Alignment with Cursor's coding workflows
3. Reinforcement Learning: Real-world developer sessions with outcome-based rewards
4. Dynamic Difficulty: Task selection gets harder as model improves

Availability

Composer 2.5 is available immediately in Cursor:

• Free tier: Limited usage
• Pro plan: Full access
• Fast/slow variants: Speed vs. cost tradeoff

Future Plans

Cursor is training an even larger model, Colossus 2, using:

• 10x more total compute
• Million H100-equivalents
• Expected to be a major leap in model capability

Note: Composer 2.5 builds on Kimi K2.5 through an authorized commercial partnership with Moonshot AI.