Bolt Offers 50x More AI Coding Capacity in Exchange for Training Data

Bolt.new, the browser-based AI development platform operated by StackBlitz, has introduced a novel arrangement with its developer community. The platform is offering substantially increased access to open-weight coding models in return for sharing anonymized training data.

On Monday, the company unveiled Forge, a research preview available to individual Pro subscribers that grants developers up to 50 times greater usage of open-weight coding models through October 14. Participation requires developers to opt in to sharing anonymized versions of their work sessions for model training purposes. This includes prompts, source code, fix traces generated during problem-solving, and conversations with the coding agent.

The shared sessions will support a collaborative effort with Arcee AI aimed at training a trillion-parameter-class open-weight model. Training is slated to commence in October, with Bolt committing to eventually release the resulting model weights to the public.

Through Forge, development work itself becomes part of the transaction. Developers receive expanded computational capacity while Bolt and Arcee obtain valuable data derived from genuine coding workflows.

Why Coding Trajectories Matter

While public code repositories contain vast quantities of source code, they typically present only finished products. A complete coding session reveals the intermediate steps: failed attempts, revisions, and the iterative path to a solution. This record of what succeeded and what failed proves invaluable as coding agents tackle increasingly complex tasks.

When agents work across large codebases, they must locate appropriate files, coordinate modifications, and handle breakdowns when issues arise. This complexity intensifies when subsequent agents inherit code written by their predecessors, which can be difficult to comprehend or adjust.

SpaceXAI demonstrated a comparable strategy last month by training Grok 4.6 using agent failure traces—the missteps, retries, and corrections that most research labs typically discard. Bolt pursues a similar strategy but sources this data from actual developer sessions rather than artificially generated scenarios.

Arcee has been investigating the same fundamental challenge. In a blog post describing NAC, its open-source agent framework, the company noted that software engineering assignments can consume tens of thousands of tokens as agents navigate code, modify files, execute tests, and resolve errors.

How Bolt Gets to 50X More Usage

Coding agents consume substantial token quantities even on single intricate assignments, making a 50-fold increase in available usage an exceptionally compelling proposition.

Forge restructures the technical foundation by executing open-weight models on Bolt's own infrastructure. The agent currently deploys GLM 5.3 Flash and GLM 5.3, with Kimi K3 and DeepSeek v4 Pro accessible as experimental alternatives.

Bolt's WebContainers technology, developed by parent company StackBlitz, provides an additional economic advantage by executing projects within an isolated sandbox inside the user's browser rather than consuming Bolt's server resources.

Forge applies this same methodology to model execution, leveraging open-weight models on dedicated hardware while developer sessions contribute to training subsequent iterations. This arrangement grants Bolt greater authority over expenses and diminishes its dependence on proprietary model APIs.

This movement toward self-hosted models is emerging across the technology sector. Nvidia's $12.9 billion acquisition proposal for Hugging Face arguably reflects an identical strategic bet, though executed at a vastly different magnitude.

Forge Scores 91% of Bolt's Top Model

Forge's open models achieved 92.2 on Bolt's proprietary Bolt Build Index, in contrast to 101.0 for its premium paid model, representing approximately 91% of the highest score. This measurement applies exclusively within Bolt's environment, so the 91% metric does not indicate how these models perform in broader contexts.

However, if Bolt can redirect more of its coding workloads toward its own models rather than purchasing proprietary alternatives, it gains greater command over expenses and consumption patterns. Simultaneously, Forge sessions furnish training material for future model iterations.

What Developers Are Giving Up

Forge mandates explicit consent, displaying a confirmation dialog whenever a developer enters the workspace. Standard and Max sessions fall outside this program, and Teams and Enterprise accounts cannot participate.

Bolt implements anonymization procedures on sessions before transmission from its systems, stripping out credentials, confidential information, and identifying details. The company validates this process using test data. Arcee obtains the resulting information under a formal data processing contract.

Developers can discontinue sharing by exiting Forge, though Bolt indicates that previously shared sessions will persist within the trained models.

The 50-fold capacity boost terminates on October 14. However, Bolt intends for Forge to continue functioning as a testing environment for open models following the preview period's conclusion.