Skip to content
Sonic
AI
Sonic
AI
Home
Discover
Ask Sonic
Projects
Use with Claude or ChatGPT
Show me around
Request source or feature
How Cursor Trained Composer on Fireworks: Distributed Infrastructure for High-Performance RL, Sonic AI
Home
/
Sequoia Capital
/
How Cursor Trained Composer on Fireworks: Distributed Infrastructure for High-Performance RL
Sequoia Capital
Notify me
•
May 26, 2026
•
45:02
Interview
How Cursor Trained Composer on Fireworks: Distributed Infrastructure for High-Performance RL
Sonya Huang
(Host)
•
Federico Cassano
(Guest)
•
Dmytro Dzhulgakov
(Guest)
Get the full transcript next time Sequoia Capital releases an episode
Summary, key quotes, top claims, and the searchable transcript — emailed automatically. No card needed.
Sign up
Executive Summary
Cursor, an AI application company, developed its own specialized agentic coding model, Composer 2, to achieve an order-of-magnitude cost reduction and performance improvement over general-purpose models.
The training process involved a two-stage approach: continual pre-training on code to build a knowledge base, followed by large-scale reinforcement learning (RL) to teach the model how to act within the Cursor environment.
In partnership with Fireworks, Cursor overcame significant infrastructure hurdles by developing a novel distributed training system that synchronized model weights across global data centers using techniques like delta compression.
This case study suggests a potential evolutionary path for AI application companies: moving from using off-the-shelf models to training highly specialized, proprietary models to optimize the trade-off between quality, speed, and cost.
Continue your research
Keep pulling the thread on Federico and Dima.
The Application-to-Model Company Pipeline
Specialization as a Competitive Advantage
Innovations in Distributed RL Infrastructure
Or ask anything across 400+ expert conversations
12
quotes
Transcript
Key Arguments
Analysis
Quotes & Entities
12
Related
Loading transcript...
Processed Jun 17, 2026
Daily intelligence brief →
yt-dlp + mlx-whisper + Gemini