Blog
2 min read

Bolt Forge: Open-Source Models, 50x More Usage — in Exchange for Your Build Data

Bolt.new launched Forge in mid-September 2026, a research-preview agent that runs only on open-weight models and gives individual Pro plans 50x more usage on them — if builders opt in to share anonymised build sessions, used with Arcee AI to train an open model. How it works, what's shared, who's excluded, and how to decide.

Bolt.new launched Bolt Forge in mid-September 2026 (most coverage dates it 14 September), as a research preview running through 14 October. It's a new agent in Bolt's picker, alongside Standard and Max, built entirely on open-weight models. (What Is Bolt.new?)

The offer

  • 50x more usage on open models for every individual Bolt Pro plan, at no extra cost, shown as a single monthly usage bar with no daily limits and a hard stop at 100%.
  • Default models reported at launch were from the GLM family, with others (such as Kimi and DeepSeek) available as experimental options. The line-up may change during the preview.

The trade

The extra usage is tied to sharing data:

  • When you switch into Forge, you get a one-tap opt-in to share that build session.
  • Shared sessions — prompts, code and fix traces — are anonymised before leaving Bolt's infrastructure, with sensitive and personal information stripped.
  • The data goes to Arcee AI, a US open-model lab, to train a large open-weight model to be released publicly.
  • You can leave at any time. Teams and Enterprise workspaces are excluded from Forge and from all data collection.

How good is it?

Bolt's own benchmark (the "Bolt Build Index") put the open models at roughly 91% of its top paid model's score. That's an internal benchmark, not an independent one — try it on your own project. For simple apps and iterations, open models are often good enough; for tricky bugs, the gap can be bigger. (Open Source Licences Explained)

How to decide

Opt in if:

  • It's a personal or learning project with nothing confidential in it.
  • You burn through usage on long iteration loops and the 50x matters.
  • You like the idea of contributing to an open model.

Don't opt in if:

"Anonymised" removes names and obvious identifiers; it can't remove the idea of your app from the code.

The broader trend

Training-data trades are spreading through AI builders: Lovable now uses Free and Pro plan data for training unless you opt out. Check every tool's data settings. (Lovable Training Data Opt-Out)

Source: The New Stack: Bolt Forge · Bolt.new announcements

EasySpawn runs Claude Code on your own server with your own Claude plan or API key — no usage-for-data trades, and your code stays in your repository. See how it works for AI-built apps or join the waitlist.

Related: Lovable vs Bolt vs v0 · Deploy a Bolt.new App · Choose an AI App Builder · Run an LLM Locally

Keep reading