Supersonic Labs Open-Weights Julia-1, a 144M-Param Decision Model Hitting 73.15% on Typed Decisions

Supersonic Labs released Julia-1, a 144.3M-parameter typed-decision model built on JHU CLSP's mmBERT-small multilingual encoder that converts (state, question, options) triples into single-shot decisions with 2-20 candidates and up to 8,192-token context. Julia-1 reports 73.15% accuracy on typed-decision benchmarks, 94% on AG News, 86% on DAIR Emotion and 71.5% on MASSIVE across 52 locales. Apache-2.0-licensed, 550 MiB checkpoint, runs on CPU or BF16 GPU — positioning it as a small-footprint Jev alternative for classification, routing and ordinal scoring.