Supersonic Labs Releases Julia 1: A 144.3M-Parameter Open Decision Model That Runs on a CPU

Supersonic Labs, a small AI lab in Brazil, has put out Julia 1: a compact model built to make choices, not to carry on a conversation. You give it context, a question, and between 2 and 20 candidate answers. It returns one pick plus a probability for each option, all in the order you supplied them.

The model is meant to handle three kinds of structured decisions through one API. It can choose a label, score an ordered set of answers, or give a yes/no probability. It also keeps caller IDs intact, which matters for routing-style use cases. What it does not do is generate free-form text.

Under the hood, Julia 1 starts from JHU CLSP’s mmBERT-small, a multilingual encoder trained on more than 1,800 languages. Supersonic Labs kept the encoder and tokenizer, added a decision head, and trained on decision-format examples. The company says this is not a fine-tuned Qwen model. The weights are on Hugging Face under Apache 2.0, and the model runs locally with Python 3.11+ on CPU or on a BF16-capable GPU. There is also an ONNX build that runs in the browser through WebGPU.

The numbers are small in the right way. Training and experiments reportedly cost about R$540, or US$104.08, in cloud GPU time. The FP32 weights take up 550.5 MiB, and the runtime supports 8,192 combined tokens, though the published benchmarks used a 1,024-token limit. Julia 2, built on the lab’s own foundation architecture, is already in development.

On benchmarks, Julia 1 looks mixed rather than magical. A September 24, 2026 evaluation gave it 73.15% on Typed Decisions, 94/100 on AG News, 86/100 on DAIR Emotion, and a rough 64/100 on Banking77. That last one is the obvious weak spot. The model also reached 71.50% macro accuracy on MASSIVE across 52 locales, with 86.25% on pt-PT and 86.75% on en-US. A separate CPU run the next day reproduced most of the numbers, including 72.55% on Typed Decisions and 60/100 on Banking77 with 3 abstentions.

Latency is where Julia 1 starts looking practical. Supersonic Labs says one decision on an Apple M4 took a median 33.15 ms. On a Samsung SM-X510 tablet through ONNX Runtime, the median was 203 ms, with 393.1 MB peak RSS. On an Intel Core i5-1235U, AG News decisions landed at 107.83 ms median, while Banking77 took 3,713.54 ms because it has to narrow 72 labels first. The company has also announced a hosted API, but it is not open yet.