Bolt Forge: 50× Open-Model Building

Introduction

Bolt.new just lit a new forge—and it’s powered entirely by open-source models. Bolt Forge is a research-preview agent that launched September 14, 2026, giving builders a sunny new lane to ship apps without leaning on closed weights. If you love WebContainers, browser-native builds, and open models that keep getting better in public, this one’s for you.

Official write-up from Donald Savard:

https://bolt.new/blog/what-is-bolt-forge

What Bolt Forge Is

Forge isn’t a rebrand of Bolt’s Standard or Max agents. It’s a new agent that runs on open-source models only, hosted on reserved hardware and StackBlitz’s WebContainers tech so your app still builds in the browser.

Out of the gate you can reach for:

GLM 5.3 Flash — the default, snappy starter
GLM 5.3 — when you want more heft
Kimi K3 and DeepSeek v4 Pro — experimental options (they burn usage faster)

Bolt’s own Bolt Build Index puts Forge’s open models at 92.2 versus 101.0 for the top paid model—about 91% of that ceiling. Not identical, but close enough that open-weight workflows feel genuinely production-curious.

The 50× Pro Promo (Through Oct 14)

Individual Pro plans get up to 50× more Forge usage at no extra cost through October 14, 2026. One monthly bar, no daily limits. When you hit 100%, Forge quietly switches you back to the Standard agent—no overage charges.

After October 14, Forge continues as Bolt’s open-model lab, but the 50× promo and the training-data collection window both end.

Pricing context from the launch: Pro is $25/mo on the yearly plan. Bolt Lite ($9/mo) is on a waitlist with first seats expected by September 21—and if you sign up by October 14, you keep that plan.

Training, Consent, and Who’s In

Training on Forge is opt-in every time you switch into Forge—you’ll see a consent screen on each switch. Standard and Max agents never train on user data. Teams and Enterprise are excluded from Forge and from data collection.

When you opt in, anonymized sessions (secrets stripped) can be shared with Arcee AI—a US open-model lab behind the Trinity family—under a DPA, to help train a trillion-parameter-class open-weight model. The first training run starts in October, and the weights are planned to publish. That’s the sunny trade: your opt-in sessions help grow the open stack you’ll eventually download.

Pro Tip

Duplicate any serious project before you flip into Forge. Treat Kimi K3 and DeepSeek v4 Pro as spicy experiments (they burn the bar faster), stick with GLM 5.3 Flash for everyday builds, and remember PDF uploads aren’t available yet. Consent is per switch—skim it each time so you always know whether that session is training-eligible.

Conclusion

Bolt Forge is a bright research preview for builders who want open models, browser-native WebContainers, and a generous Pro usage bump through mid-October. Start on GLM 5.3 Flash, keep serious work duplicated, and decide consciously whether each Forge session should help train the next open-weight giant.