Bolt Offers Developers 50x More AI Compute—if They Share Their Coding Sessions

StackBlitz's Bolt.new has launched Forge, a research initiative that dramatically expands model usage for Pro subscribers willing to contribute their development activity to model training. Through mid-October, participants gain access to up to 50 times more compute from open-weight coding models—a substantial increase for developers working on complex tasks. In return, they must allow StackBlitz and partner Arcee AI to collect anonymized records of their coding sessions, including prompts, source code, debugging traces, and interactions with the AI agent.

The collected sessions will feed a trillion-parameter-scale open-weight model, with the first training cycle beginning in October. StackBlitz has committed to releasing the resulting model weights publicly once training concludes.

Why Developers' Trial-and-Error Data Matters#

Training data sourced from public repositories reveals mostly finished code. What's missing is the messy middle—the failed attempts, corrections, and debugging steps that developers encounter as they solve real problems. These trajectories are particularly valuable for teaching AI agents to handle complex, multi-step coding tasks.

When AI agents inherit code written by other agents or navigate large codebases, they face unique challenges: locating the right files, coordinating changes across multiple files, and recovering from failures. Recording these interactions helps models learn resilience and adaptation in ways synthetic training data often cannot capture.

SpaceX AI employed a similar strategy last month by training Grok 4.6 on agent failure traces—capturing the missteps and corrections that most research typically discards. Bolt's approach mirrors this philosophy but draws real-world data from actual developer workflows rather than controlled experiments.

How Bolt Achieves 50x Increased Usage#

Achieving a 50-fold jump in available compute requires rethinking infrastructure. Forge runs open-weight models on Bolt's own servers rather than relying on proprietary model APIs, giving the company direct control over costs and capacity. The current selection includes GLM 5.3 Flash and GLM 5.3, with experimental options for Kimi K3 and DeepSeek v4 Pro.

StackBlitz's WebContainers technology—which executes projects in an isolated browser environment rather than on remote servers—provides additional cost savings. By extending this philosophy to model inference, Forge reduces operational expenses while offering users substantially more coding power.

This shift toward self-hosted infrastructure reflects a broader industry trend. Nvidia's pursuit of Hugging Face signals the same strategic move at a vastly different scale: building in-house model capabilities rather than remaining dependent on external APIs.