AI 程式開發工具
Bolt Forge Trades 50× Agent Credits for Secret-Scrubbed Code-Repair Trajectories
Bolt.new’s research preview runs its website-building agent on open-weight models from GLM, Kimi, and DeepSeek, while providing Arcee AI with prompts, code, and error-repair sequences from users who explicitly opt in for training. The program clearly illustrates a “compute for data” model, but its self-reported 91% benchmark result and secret-scrubbing effectiveness still require external validation.
Bolt.new launched Forge on September 14. It is not a new standalone model, but a third execution mode for the platform’s agent. It uses GLM 5.3 Flash by default, with GLM 5.3 and experimental access to Kimi K3 and DeepSeek v4 Pro also available. Until October 14, individual Pro users can receive up to 50× Forge credits in exchange for explicitly consenting to share the relevant build session each time they switch to Forge.
The data includes more than the final code. Bolt says it will collect prompts, edits, runtime errors, retries, and repair trajectories. It will first remove secrets and personal information and anonymize the data on its own infrastructure, then provide it to Arcee AI under a data processing agreement. Compared with public repositories, these step-by-step trajectories reveal more about how an agent progresses from failure to a working version. They will be used in the first round of training beginning in October, with the two companies aiming to train an open-weight model at trillion-parameter scale.
The cost structure is also technically significant. Forge uses inference hardware reserved by Bolt, while projects continue to run inside browser-based WebContainers, avoiding the need to rent a separate server for every build. Models consume credits at different rates: Kimi K3 and DeepSeek v4 Pro exhaust the quota faster than the GLM options. Once Forge credits run out, the platform reverts to Standard without automatically generating overage charges. Teams and Enterprise workspaces are entirely excluded from the data program.
Bolt’s Build Index shows the Forge model lineup scoring 92.2, about 91% of the 101.0 score achieved by the top paid model. However, this is the platform’s own benchmark, and it has not disclosed the full set of tasks, variance, or per-model results. The figures therefore cannot be interpreted directly as a nine-percentage-point difference in production quality. A more important risk concerns secret scrubbing: even if the filters are validated with seeded test data, code may still expose organizational information through business logic, data structures, or comments. Moreover, data that has already entered training cannot subsequently be removed from model weights. Key issues to watch include whether Bolt discloses its data-retention period, model licensing, training-set audits, and independent benchmarks.