A tiny AI lab in Brazil just dropped a decision model that runs on your laptop, costs nothing to use, and was trained for a total of US$104. That last detail is the whole story.
What's happening
Julia 1, released September 26th by Supersonic Labs, is a 144-million-parameter open-weight model that does exactly one thing: it makes choices. Give it a question, a set of options, and some context, and it picks the right answer. No text generation, no long responses. Just: "Option B."
That might sound limited. It's actually incredibly useful. The tasks a small business repeats every day, flagging urgent emails, routing support tickets, sorting customer feedback by type, are all routing and classification problems. Julia 1 solves those in 33 to 107 milliseconds on whatever CPU you already own.
The weights are published on Hugging Face under Apache 2.0, so you download once and run it forever, offline, for free. It also runs in a browser via WebGPU. The team trained the whole thing on a cloud budget of $104. That is not a typo. That is what it looks like when a scrappy independent lab decides to actually build something.
Try this this week
- Route your support inbox: Feed Julia 1 a snippet of each incoming email and a short list of categories (billing, appointment, complaint, general) and let it tag them automatically. Wire the labels into a tool like n8n to sort into folders or ping the right Slack channel.
- Classify customer feedback: Paste in a review and ask "positive, neutral, or negative." Instant sentiment without sending anything to a cloud API.
- Build a yes/no filter: Ask Julia 1 whether a message crosses a threshold ("Is this message urgent?") and put a lightweight triage layer in front of your team's inbox before it gets noisy.
- Get started: Download the weights from the Hugging Face repo, follow the Python quickstart in the Supersonic Labs docs, and run your first classification in under ten minutes on Python 3.11+.
- One limit to know: Julia 1 struggles when options are numerous and similar-sounding. Keep your category lists small (under 20) and make each label distinct.
The bigger picture
The best AI tool for a small team is usually a specialized one. Julia 1 is not trying to replace a general assistant. It is trying to do one job well: classify and route, without a monthly subscription, without sending data anywhere, and without needing a server. That is the smallest-working-thing philosophy in action, and it is exactly the kind of building block a two-person shop can actually operate.