Back to rankings

Bolt Forge: An Open-Weight Agent With 50X Usage

BoltView original changelog

Bolt launched Bolt Forge, a new agent that runs exclusively on open-weight models (GLM 5.3 Flash, GLM 5.3, Kimi K3, and DeepSeek v4 Pro) alongside its existing Standard and Max agents. Every individual Pro plan gets up to 50X more usage through Forge at no extra cost through October 14, 2026, in exchange for opting in to share anonymized build sessions used to train open models with partner Arcee AI. Bolt says Forge's open models reach 91% of its top paid model's score on its internal Bolt Build Index benchmark.

Key Takeaways

  • Bolt launched Bolt Forge, a third agent option built entirely on open-weight models (GLM 5.3 Flash by default, plus GLM 5.3, Kimi K3, and DeepSeek v4 Pro).
  • Individual Pro subscribers get up to 50X more usage through Forge at no extra cost through October 14, 2026.
  • Using Forge means opting in to share anonymized build sessions for training open models with partner Arcee AI; Standard and Max never train on user data.
  • On Bolt's own Bolt Build Index benchmark, Forge's open models scored 92.2 versus 101.0 for Bolt's top paid model, roughly 91% capability parity.
  • Forge is experimental inside Bolt: the company recommends duplicating serious projects before using it, and it cannot yet accept PDF uploads.
  • Teams and Enterprise workspaces are excluded from Forge and from all associated data collection.

Bolt Introduces Bolt Forge, an Open-Weight Agent

Bolt launched Bolt Forge, a new agent inside Bolt.new that runs exclusively on open-weight AI models rather than the proprietary models that power Bolt's existing Standard and Max agents. Forge enters as a research preview and sits alongside Standard and Max in the agent picker as a third, distinct option.

An Open-Model Lineup

Forge defaults to GLM 5.3 Flash, with GLM 5.3 available alongside it, and Kimi K3 and DeepSeek v4 Pro offered as experimental alternatives. Bolt says the lineup will keep evolving, with new open-weight models landing in Forge first as they become available. On Bolt's own benchmark, the Bolt Build Index, Forge's open models scored 92.2 against 101.0 for Bolt's top paid model, which Bolt frames as 91% capability parity at a fraction of the cost.

50X More Usage, in Exchange for Opt-In Training Data

The incentive for switching agents is usage. Every individual Pro plan gets up to 50X more usage through Forge at no extra cost, through October 14, 2026, drawn from a single monthly allocation with no daily limits. The trade is that Forge sessions are used to train new open-weight models unless the user declines: a one-tap consent screen appears every time someone switches into Forge, and builds made in Standard or Max are never used for training. Bolt says shared sessions are anonymized and stripped of secrets and personal data before being sent to training partner Arcee AI, a US open-model lab, under a signed data processing agreement. Teams and Enterprise workspaces are excluded from Forge entirely.

Why Bolt Built It

Bolt frames Forge as a response to how usage-based pricing discourages the exploratory, trial-and-error phase of building, where users historically rationed prompts even during early drafts. Forge's larger allocation is meant to remove that friction for brainstorming and prototyping, while keeping production-grade work on Standard or Max. Bolt also points to a steep drop in AI inference costs over the past 18 months as what makes an allocation this large financially viable.

Access and Limitations

Forge is available to individual Pro subscribers with no separate signup, and a waitlisted lower-cost "Bolt Lite" tier opens its first seats later in September. Bolt cautions that Forge's models are still experimental inside the product: it recommends duplicating any serious project before switching it into Forge, notes that Kimi K3 and DeepSeek v4 Pro consume usage faster than the GLM models, and says Forge cannot yet accept PDF uploads. After the October 14 preview window closes, Forge continues as an ongoing open-weight agent inside Bolt, though the 50X allocation is tied to the preview period.