It’s time to accelerate open weight models to the frontier.

And bring abundant tokens to all.

Today we launch Forge in partnership with Arcee, Microsoft, Vercel, Fireworks, and DigitalOcean: https://t.co/EJ5b8Mf9Wc

Accelerating Open Weight Models to the Frontier

This weekend the leaders of the largest closed AI labs agreed the industry should slow down.

I want to talk about the other direction: acceleration to the frontier.

The gap is closing

For two years, the story of open weight models was “impressive, but eighteen months behind.” That number has collapsed. GLM, DeepSeek, and Kimi now build and ship real software. The delta to the frontier is now measured in months, and every release cycle shrinks it.

In parallel, frontier lab products are rationing tokens. Usage caps, throttles, plans that run dry mid-project, API pricing at 80%+ margin. The industry built tools that made builders 10x faster and then metered the fuel.

That is backwards. It is time to fulfill the promise of AI to the public and pass along the abundance, and see what people build when they stop counting tokens.

Forge

Today we are launching Forge, a research preview built in collaboration with @Arcee_ai, @Microsoft, @Vercel, @FireworksAI_HQ, and @DigitalOcean.

What it is: up to 50x usage on the best open weight models in production anywhere, included for free on every Bolt Pro plan for the next 30 days. It is also the heart of a new low-cost $9/mo plan (more below).

To start, GLM 5.3, Kimi K3 and DeepSeek v4 are available in the model picker. More models will be continually added as we secure infrastructure capacity. Nothing to configure, nothing to migrate. Pick a model and build.

How abundance actually works

Abundant tokens do not come from subsidies. They come from structure.

We spent the last few months locking in the largest amount of GPUs at the lowest possible prices with our infrastructure partners, and the last weekend getting the final pieces over the line.

Our goal is for Forge to run open weight models at 50% below raw open model inference cost. That is not a promotional discount. It is what becomes possible when the inference is funded by something more valuable than margin: the research program underneath it.

Here is the deal, stated plainly: Opt in to Forge and your sessions, fully anonymized, get used as training data to forge frontier open weight models by Western AI labs.

To start, we are helping Arcee, one of the leading American open weight model labs, train their first ever trillion plus parameter open weight model.

**In exchange, you get up to 50x usage on the best open models available. **You get abundance. American open source gets the data to compete with the frontier. Everyone is paid in the currency they actually need.

Yes, the best open models today are Chinese

While all models on Forge are hosted & served from Western countries, the model weights for GLM/Deepseek/Kimi themselves come from Chinese labs.

Someone is going to say it, so let me say it first: “Bruh, you’re just distilling Chinese models”.

Great point chap, let’s talk about it, because students of history know this is the oldest and most American play in the book.

In 1789, a young millworker named Samuel Slater memorized the designs of Britain’s textile mills and sailed to Rhode Island, because exporting them on paper was illegal. England called him a traitor. America calls him the father of the American industrial revolution. Every industrial catch-up in history starts by borrowing the leader’s tools. Using the best available technology to build your own is not a contradiction of the goal. It is the playbook. This is more American than apple pie- in particular because apple pie came later.

And let’s not forget that every frontier model alive, American ones included, learned from the outputs of the models that came before it. And before that, the entire corpus of the internet, and every book ever written in human history.

And when the open weight models Forge helps create get released, it will further accelerate the ability for more open models to train off of them.

So yes, the best open models today are Chinese. That is not a gotcha. That is the problem statement. Forge is the fix: builders of the free world, building on the best available models, generating the data that trains the next generation of open weights on our side of the frontier.

Bolt Lite

One more thing. Abundance should be accessible at a price point anyone can pay.

Today we are opening the waitlist for Bolt Lite: $9/month, 50x usage on Forge models, with first admits starting this week. Sign-ups close October 14. Lock in $9 and it renews at $9 for as long as you keep it.

The other direction

The frontier will set its own pace. That debate matters and the relevant people are having it.

But below the frontier there is a different project, and it does not need anyone’s permission: take the models that already exist, make them abundant, and put them in the hands of everyone with something to build.

That project starts today.

If you are a Western open weights lab or infrastructure provider and want to help accelerate open weights to the frontier, join us.

DMs open @ericsimons as is my inbox [email protected]