Julia 1 is a decision model released on September 26, 2026 by Supersonic Labs, a small Brazilian lab: just 144.3 million parameters, designed for text classification and routing on CPUs. The weights file is 550.5 MiB; code and weights are both on Hugging Face, open-sourced under Apache 2.0. It doesn't generate text — it only chooses. That's the first thing to get straight about it.
Get the concept right first: it picks answers, it doesn't write them
Julia 1 takes "context + question + a set of candidate answers" as input and outputs one pick from the candidates. Its jobs are specific: text classification, request routing, ordered scoring, and yes/no judgments. It's built on the multilingual mmBERT-small encoder — note, it is not a large generative model, so don't expect it to write anything.
How to read the official numbers: small pilots, not production evidence
The lab published a few small-scale pilot results: on Banking77, 100 samples with 72 labels first narrowed down via a shortlist, 64 answered correctly, with the lab's own comparison figure at 87%; news classification: 94 out of 100 correct across 4 labels; sentiment classification: 86 out of 100 correct across 6 labels. To be clear: these are company-reported small-sample pilots, not production evidence, and figures measured under different protocols can't be directly compared. The on-device numbers are more down-to-earth: it has run on an Apple M4, an Intel i5-1235U laptop, and a Samsung tablet, doing about 5 decisions per second on CPU on the tablet — the lab says its available acceleration path fell back to CPU, so treat that as a deployment demo. The lab also disclosed that cloud GPU training and experiments cost about 540 Brazilian reais in total, roughly $104.
Limits and pitfalls: read these before you try it
The number to be most wary of is the 64%-vs-87% comparison on Banking77: the lab's protocol measured accuracy after narrowing the labels with a shortlist, so the two figures weren't tested under equal conditions — comparing them directly misleads. Anyone evaluating it should re-test on their own label taxonomy instead of taking the launch copy's numbers at face value. And the positioning bears repeating: it's a decision model, not a generative one — using it for generation or reasoning is the wrong tool for the job. Small-pilot results shouldn't be extrapolated straight to production either.
The agent era still needs a cheap "routing layer"
The agent era needs more than bigger generative models — it needs a cheap "routing layer": if every classification and every routing decision requires a call to a cloud-scale model, neither cost nor latency will hold up. Small models like Julia 1 exist for exactly this — keeping high-frequency, low-risk judgments local, solved by small models. Supersonic Labs states its positioning plainly: building "practical AI that runs on hardware people already have." A model trained for $104, making 5 decisions per second on a tablet — the point of this story was never leaderboard-chasing, but getting AI onto devices ordinary people already own.
