Case Study: Cursor achieves frontier coding performance with Fireworks AI

Cursor boosts coding model performance with Fireworks AI, cutting inference cost 6–10x

Cursor, a software development company, sought to build a highly specialized AI coding model optimized solely for its own integrated development environment. The challenge was to move beyond general-purpose foundation models and create an agent that could improve through real developer workflows, which required scaling large-scale reinforcement learning (RL) training across distributed infrastructure without building a dedicated internal inference stack. They partnered with vendor Fireworks AI to address this.

Fireworks AI provided the high-performance inference and distributed rollout infrastructure that made Cursor's RL training practical. The solution enabled distributed RL across multiple global clusters with compressed weight synchronization, reusing production inference to accelerate training. The result was Cursor Composer 2, which achieved frontier-level coding performance on key benchmarks while reducing inference costs by 6–10x compared to similar models and significantly improving reliability in long-horizon tasks.

Cursor