Supersonic Labs Ships Julia-1 Semantic Router

Supersonic Labs has launched Julia-1, a lightweight decision model designed for text classification and semantic routing. Released under an Apache 2.0 license, the 144.3-million-parameter model is the first public release from the company's proprietary training system. Julia-1 is built on top of mmBERT-small, a multilingual ModernBERT encoder developed by the Johns Hopkins University Center for Language and Speech Processing (JHU CLSP). Supersonic Labs added a custom classification layer to score candidate answers, training the combined system on decision-format examples.

At its core, Julia-1 functions by accepting a state, a question, and between 2 and 20 candidate options, then scoring those options in their supplied order. Because candidate meanings are included directly in each request, developers can dynamically alter classification labels without needing to retrain the model or deploy a new output head. The model supports three distinct typed modes through a single API: choice, ordered score, and Boolean decisions.

In performance evaluations, Julia-1 achieved a score of 73.15% on typed decisions. It also recorded 94% accuracy on the AG News dataset, 86% on the Emotion benchmark, and 71.5% on the MASSIVE dataset across 52 locales. However, the model struggles with longer label lists, scoring just 64 out of 100 on a Banking77 pilot compared to the reference score of 87 out of 100.

For deployment, Julia-1 is highly versatile and can run locally on a standard laptop CPU. Supersonic Labs has also provided an ONNX and WebGPU build that enables the model to run directly in a web browser. This in-browser setup delivers a latency of 75 milliseconds per decision while maintaining a 100 out of 100 parity with PyTorch execution. The release includes the model weights and inference code, though the training pipeline remains proprietary.