Supersonic Labs releases Julia 1: a 144.3M-parameter open decision model that runs on CPU

Brazilian lab Supersonic Labs has released Julia 1, a compact 144.3M-parameter decision model that runs on a plain CPU. Weights are on Hugging Face under Apache 2.0, with an ONNX build for browsers via WebGPU. The model picks one of 2–20 candidate answers and returns probabilities; it does not generate text.

• 144.3M parameters, FP32 weights take 550.5 MiB, training cost about $104
• One API covers choice, ordered score and yes/no decisions over 2–20 options
• Median latency 33.15 ms per decision on Apple M4, 203 ms on a Samsung tablet
• Clear failure on Banking77: 64/100 vs 87% for the Jev baseline