In early August, Open Weights and American AI Leadership closed to new signatures. More than 270 companies and organizations signed it, from Microsoft, Google, and NVIDIA to Hugging Face, Red Hat, and the Linux Foundation.
NimbleBrain is a proud signatory.
The letter went out July 24. It argues that open-weight models, the ones anyone can download, run, and adapt, are central to how America builds with AI. We signed because of who we build for.
We do some of our work in the most remote places in the country, Hawaiʻi among them. Shipping, broadband, and software all reach the islands later, cost more, and come on someone else’s terms. Most of the businesses I talk to there will never train a model, and open weights give them the same foundation as a company in San Francisco, on hardware they control.
A signature should come with commitments, so I’m putting ours in writing.
Frontier models earn their place
We run frontier models at NimbleBrain every day, and they still surprise me. When a problem needs the best reasoning available, or a mistake costs more than the tokens, I reach for a frontier model. I want the labs pushing the ceiling to keep pushing.
Most of the work AI will do sits well below that ceiling. Think summarizing a contract, routing a support ticket, or matching an invoice to a purchase order. That work needs a model that’s good enough, cheap enough to run all day, and yours. Call it the floor. It’s where most businesses will do most of their AI work.
Who gets to stand on the floor
The letter says “Our AI leadership will be judged not by one frontier AI model.” I agree, and I’d name the measure: how many people get to build with this technology, and on whose terms.
We’ve run this experiment before. In the mid-1930s, about nine in ten American city homes had electricity and about one farm in ten did. Utilities saw no profit in stringing wire out to farms, so rural families waited until the Rural Electrification Act of 1936 forced the issue. A generation fell behind for reasons that had nothing to do with how hard they worked.
AI can split the same way. If capable models belong to a handful of companies and everyone else rents access, the renters have no say when prices or terms change. Left long enough, that gap hardens into economic stratification, and stratification breeds strife.
Open weights cut the other way. A school district, a county government, or a five-person startup gets the same foundation as a Fortune 500. They run it where they choose, tune it to their work, and keep their data. No model provider can raise the price on them or retire the model out from under them. And as running a model gets cheaper every year, owning it matters more than the price of a token.
What we’re committing to
Four things, and you can hold us to them.
Our runtime is model-agnostic. The NimbleBrain runtime runs open-weight and frontier models, and you pick. Swapping one for another is a configuration change.
You can run open models inside your own boundary. Your models run on your infrastructure, and your data stays in your environment.
NimbleBrain is open source. The NimbleBrain runtime is Apache 2.0. You can read every line, run it on your own infrastructure, and keep running it whether or not you ever pay us.
We recommend open models when they do the job. If a client’s task runs well on an open model, that’s what we recommend, even when a frontier model scores a little higher. Good enough and yours beats slightly better and rented. When the work needs the ceiling, we’ll tell you.
Put an open model to work
I challenge you to pick one task your team runs every week and run it on an open model alongside the one you use now. Compare accuracy, cost, and how much review the output needs. I bet the floor sits higher than you think. If you want help running it inside your own walls, come talk to us.
Share this post