Supersonic Labs, a small AI lab from Brazil, has launched Julia 1. It’s a compact determination mannequin, not a chatbot. You cross it context, a query, and a pair of to twenty candidate solutions. It picks one and returns a chance for each choice. The mannequin has 144.3M parameters and runs on a plain CPU.
Is it deployable? Sure. The weights are on Hugging Face beneath Apache 2.0 and run domestically with Python 3.11+ on CPU or a BF16-capable GPU. An ONNX build additionally runs within the browser by way of WebGPU. A hosted API is introduced however not open but.
What Julia 1 Does
Julia 1 handles three determination varieties by means of one API:
- alternative: choose one label from 2 to twenty described choices (classification, routing).
- rating: return the anticipated index on an ordered rubric, reminiscent of low, medium, excessive.
- noul: return the chance {that a} yes-or-no assertion is true.
Outcomes come again within the caller’s choice order with full softmax chances. Caller IDs reminiscent of billing are returned unchanged. The mannequin doesn’t generate textual content.
Structure and Coaching Funds
Julia 1 begins from JHU CLSP’s mmBERT-small, a 140M-parameter multilingual ModernBERT encoder educated on 1,800+ languages. Supersonic Labs saved the encoder and tokenizer, added a choice head, and educated on decision-format examples. The lab states Julia 1 is not a fine-tuned Qwen mannequin. The runtime helps 8,192 mixed tokens, however printed benchmarks used a 1,024-token restrict.
Whole cloud GPU spend for coaching and experiments was about R$540 (US$104.08). The FP32 weights occupy 550.5 MiB. The non-public coaching pipeline will not be launched. Julia 2, with the lab’s personal basis structure, is in growth.
Benchmark Outcomes
The September 24, 2026 analysis ran on H200 BF16 with strict encoding. The comparability baseline is TypeSafe’s Jev, utilizing reference values from the Jev benchmark protocol, not a brand new Jev run.
- Typed Decisions: 73.15% (1,463/2,000) vs 72.70% reference.
- AG Information, 4 labels: 94/100 vs 91% reference.
- DAIR Emotion, 6 labels: 86/100 vs 48% reference.
- Banking77, 72 labels: 64/100 vs 87% reference. That is the clear failure.
- MASSIVE, 18 situations: 71.50% macro accuracy throughout 52 locales; 86.25% pt-PT, 86.75% en-US.
The classification pilots use solely 100 examples every. A September 25 CPU run reproduced most numbers: 72.55% on Typed Choices and 60/100 on Banking77 with 3 abstentions.
On-Machine Latency
The lab printed per-device measurements. On an Apple M4, one determination per name took a 33.15 ms median. On a Samsung SM-X510 pill by way of ONNX Runtime, the median was 203 ms with 393.1 MB peak RSS. On an Intel Core i5-1235U, AG Information selections took a 107.83 ms median. Banking77 took 3,713.54 ms as a result of it narrows 72 labels first.
On X, @supersonicai claims Julia 1 classifies 5x quicker than Jev on an i5 laptop computer. Deal with that rigorously. The Jev pilot measured Jev as a hosted service referred to as from France, so latencies are usually not like-for-like.
Interactive Explainer
‘;doc.getElementById(‘dots’).innerHTML=h;
doc.getElementById(‘rSteps’).innerHTML=”;doc.getElementById(‘rOut’).textContent=n‘+r[0]+’
‘+r[3]+’ · ‘+(d>0?’+’:”)+d+’ pp
Julia 1: ‘+fm(r[1])+’%
Jev ref: ‘+fm(r[2])+’%
