Bolt.new launches Forge to widen who can build with open models

Bolt.new launched Bolt Forge, a research-preview agent that runs only on open-source models and gives every individual Pro plan up to 50 times more usage through October 14 — if the builder opts in to share anonymized sessions that help train new open-weight models with Arcee AI.

Paid coding agents still ration who can draft. A same-day company primary that trades extra open-model usage for opted-in build sessions is a software-desk read on whether open weights get trained on real software work — not a claim the planned trillion-parameter model exists yet.

Bolt.new, the AI app builder built by StackBlitz, said via Business Wire on 14 Sep 2026 that it launched Bolt Forge, a research preview meant to widen who gets to build software on open-source models. An open-source model here means AI software whose weights — the trained numbers that make it work — can be released for other people to download and run. That company release is the filing event. These are company claims. This desk did not run Forge.

Forge, as Bolt.new tells it, is a new agent — a helper that writes and edits an app from a prompt — inside Bolt.new. It sits next to the Standard and Max agents already on the product. It runs only on open-source models. Premium modes stay private and untouched: Bolt says it never trains on anything built outside Forge. File that split as the company’s. This desk did not audit the training pipeline.

The lineup, as the same release and the company blog name it: GLM 5.3 Flash; GLM 5.3 alongside it; and Kimi K3 and DeepSeek v4 Pro as experimental options. The company blog says GLM 5.3 Flash is the default. File the names as Bolt.new’s. This desk did not score those models.

The trade, as Bolt.new and Arcee AI tell it: builders who opt in share anonymized sessions — prompts, code, and fix traces with secrets and personal information stripped — and that data goes to Arcee AI, a U.S. open-weight lab behind the Trinity model family. Bolt.new says it has partnered with Arcee to help train a trillion-parameter-class open-weight model. A parameter is one of the trained numbers inside a model; a trillion-parameter-class model is a very large one, not a size this desk measured. A signed data processing agreement, the companies say, governs every transfer. File the trade and the size class as theirs. This is not a claim the trillion-parameter model is already trained or downloadable.

Eric Simons, co-founder and CEO of Bolt.new, quoted on the release: “With Forge, you forge your own products and businesses.” And: “by building, you're helping the world forge capable open-weight models. A year ago, a lot of these builders didn't have the purchasing power to do this at all. Arming them is the whole point.” File the quotes as his. They are not an independent product review.

Mark McQuade, co-founder and CEO of Arcee AI, quoted on the same page: “When a model's weights are open, capability stops being something you rent.” He said models trained on Forge are released to the public, so anyone can download and run them, including the builders whose work made them better. “Widening who gets to build software starts with widening who owns the models, and open-source models belong to everyone.” File the quotes as his. This is not a download link for a finished trillion-parameter model.

Consent, as Bolt.new describes it: a one-tap opt-in fires every time a builder switches into Forge. Decline it, and nothing is shared. Builders can leave by switching back to Standard or Max; data already used in training stays in the models it helped train. Teams and Enterprise workspaces are excluded from Forge and from all data collection. File the consent picture as the company’s. This desk did not watch the consent screen.

Usage, as the company tells it: every individual Pro plan includes up to 50 times more Forge usage at no extra cost through October 14, 2026. Usage shows as one monthly bar with no daily limits and a hard stop at 100 percent — when the bar fills, Bolt says it switches the user back to Standard rather than charging an overage. Forge usage is separate from Standard and Max. File 50×, the bar, and the hard stop as Bolt.new’s. This desk did not meter an account.

How good the open models are, as the company blog has it: Forge’s open models score 92.2 on the Bolt Build Index against 101.0 for the top paid model in Bolt.new — 91 percent of that top score. The Bolt Build Index is Bolt.new’s own test of how well a model finishes real Bolt projects. File 92.2, 101.0, and 91 percent as company-benchmark figures. This desk did not rerun the index, and it is not treating those numbers as an independent bake-off.

The research preview runs from September 14 to October 14, 2026, the release says. Bolt Forge is live today for individual Pro subscribers, with more at bolt.new/forge. After October 14, the company blog says Forge continues as an open-model lab, while the 50× allocation and the first training-data window with Arcee both close that day. File the calendar as Bolt.new’s. Do not upgrade this to a generally available product for Teams or Enterprise, or to a claim the first training run has already finished.

PRIMARY here: Bolt.new’s 14 Sep 2026 Business Wire release — Tier A company source, not a second-hand wire — plus the same-day MarketMinute reprint of that same originating text, the company Forge explainer, and Arcee AI’s same-day partner note. The 50× allocation, the trillion-parameter-class target, and the 92.2 / 101.0 Build Index scores are company-attributed. NOT claimed: that the trillion-parameter model is already trained or public, independently measured quality, independently verified anonymization, that this desk ran Forge, a Teams or Enterprise launch, a stock tip, or investment advice. Distinct from the already-filed sourcegraph-agentic-batch, apple-ios-27-siri-ai, cognition-swe-2, perplexity-portable-windows-rtx, and sokin-mcp-payments.