Eric Simons' AI app builder is feeding anonymized coding sessions to Arcee AI for a trillion-parameter-class open model.
By [Ryan Merket](/author/ryan-merket)
· Published
Primary source: [X / bolt.new](https://x.com/boltdotnew/status/2099529089430229059)
Why it matters #
Bolt is turning discounted inference into a data-acquisition engine, using real coding sessions to shape open models around its product instead of depending entirely on frontier labs.
Eric Simons (@EricSimons), co-founder and CEO of Bolt.new, launched Bolt Forge on September 14th, giving paid individual users up to 50 times their usual AI-building allocation in exchange for opting into a training-data program with Arcee AI.
https://x.com/boltdotnew/status/2099529089430229059 The research preview runs through October 14th. Forge appears beside Bolt's Standard and Max agents and uses open-weight models including GLM 5.3 Flash, GLM 5.3, Kimi K3 and DeepSeek V4 Pro. Bolt says Forge runs on reserved hardware, replacing per-request model-provider fees with a more predictable infrastructure bill.
That cost structure lets Simons make an aggressive usage offer. It also gives Bolt and Arcee something model developers cannot readily scrape from the public web: complete records of people attempting to build software, encountering errors and asking an agent to repair them.
Simons built Bolt on years of browser infrastructure work with StackBlitz co-founder Albert Pai. The childhood friends taught themselves web development before founding StackBlitz in 2017, according to Simons' account of their early work. Their WebContainers runtime executes full development environments inside a browser rather than renting a remote server for each project. Forge extends that cost advantage from software execution to model inference and training data.
The extra usage has a price
Forge users must consent each time they switch into the agent. Bolt says shared records include prompts, generated code and the traces created while the agent fixes errors. Bolt says it strips secrets, personal information and other sensitive data before sending sessions to Arcee AI, with transfers governed by a data-processing agreement.
Users can stop future collection by returning to Standard or Max. Data already used for training cannot be withdrawn from the resulting models. Bolt's Teams and Enterprise workspaces are excluded from Forge and its collection program, limiting the preview to individual Pro accounts and the new Lite subscription.
The distinction matters because the headline offer is not unlimited free computing. Forge has a separate monthly allocation with no daily limit, according to Bolt's product explanation. Usage stops at 100%, at which point Bolt moves the user back to Standard rather than imposing an overage charge. The "zero usage charges" pitch therefore means no metered overage inside Forge, while access still requires an eligible paid plan.
Individual Pro plans, priced at $25 per month when billed annually, receive the expanded allocation through October 14th. Bolt also introduced Bolt Lite, a $9 monthly plan built around Forge for students and side projects. Access codes are rolling out in waves, and sign-ups close on October 14th. Customers admitted before the deadline can keep renewing Lite at $9 per month, according to Bolt.
Bolt becomes part model lab
Arcee will use opted-in sessions from the preview in a training run scheduled to begin in October. Bolt says the partnership is working toward a trillion-parameter-class open-weight model whose weights will be published after training.
Arcee already develops the Trinity family of downloadable models under permissive licenses. Its existing work provides the training infrastructure; Bolt supplies a stream of task data tied directly to browser-based software development. The arrangement moves Bolt beyond choosing models from outside laboratories and gives Simons a route to influence models around the specific workflows his product needs.
The initial Forge lineup reflects how quickly open models have expanded into coding and agent work. Kimi K3 is a 2.8 trillion-parameter mixture-of-experts model with open weights and a one-million-token context window. GLM 5.3 Flash uses 320 billion total parameters with 18 billion active during inference. DeepSeek released V4 Pro in August as a model aimed at agent and coding workloads.
Bolt positions those models as cheaper substitutes for proprietary systems, though its supporting performance number comes from its own benchmark. Forge scored 92.2 on the Bolt Build Index, compared with 101 for the highest-scoring paid model, Claude Opus 5, according to Bolt. The benchmark methodology has not been independently validated, and Bolt advises users to duplicate important projects before using Forge and to keep complex production work in Standard or Max.
That warning defines the launch more accurately than the 50x figure. Forge gives Bolt users a large sandbox for drafts and experiments while converting their work into a proprietary training corpus for an openly released model. Simons is subsidizing inference for one month to find out whether that exchange produces enough useful data, lower costs and model improvement to become a permanent part of Bolt.