Supersonic Labs Launches Julia 1, a CPU-Friendly 144.3M-Parameter Decision Model
Supersonic Labs released Julia 1, a 144.3M-parameter decision model based on mmBERT-small that chooses among 2 to 20 options and returns probabilities. It runs on CPU, is available under Apache 2.0, and supports choice, score, and yes-or-no decision modes. Benchmarks showed improvements on three of four pilots but weaker results on the 72-label Banking77 test.
Julia 1 comes from a small Brazilian lab and is framed as a decision component, not a conversational system. It builds on a multilingual encoder from JHU CLSP, adds a decision head, and offers choice, score, and yes/no modes through one interface. Results keep the order of options supplied by the caller and include softmax probabilities; caller identifiers pass through unchanged, and the model produces no text.
The weights are Apache 2.0 and can run locally on CPU or BF16 GPU, with an ONNX/WebGPU browser path. Training and experiments reportedly cost about US$104. Benchmarks favored it on Typed Decisions, AG News, and DAIR Emotion, but Banking77 with 72 labels lagged. A CPU rerun largely matched those results.
Open, CPU-capable decision models like Julia 1 could make automated routing, triage, and simple classification more accessible to small teams, students, and privacy-sensitive deployments that cannot rely on hosted APIs. Because it runs locally and returns probabilities, developers may embed it in edge devices or multilingual services at low cost. Its weaker 72-label result, however, may keep it out of complex, high-stakes tasks until validated further. The main beneficiaries could be resource-constrained builders and users wanting faster, local decisions; limitations may affect accuracy-sensitive domains.