Supersonic Labs Ships Julia-1, a Tiny 144M Router That Classifies Without Retraining
Supersonic Labs released Julia-1, a 144M-parameter open decision model that turns a state, question, and options into one typed choice, running in-browser on WebGPU.
- Supersonic Labs released Julia-1, a 144.3M-parameter Apache 2.0 decision model for classification and routing.
- Fine-tuned from mmBERT-small, it takes state + question + 2-20 options and picks one.
- Supports three typed modes: choice, ordered score, and Boolean (noul) decisions through one API.
- Scores 73.15% on typed decisions, 94% AG News, 86% Emotion, 71.5% MASSIVE across 52 locales.
- Weak on long label lists: 64/100 on Banking77 pilot versus 87/100 reference.
- ONNX/WebGPU build runs 75 ms per decision in-browser with 100/100 parity to PyTorch.
Supersonic Labs has released Julia-1, a compact model that selects an answer from a supplied set of candidates. Given context, a question, and between two and 20 options, it can classify text, route requests, score ordered rubrics, or make Boolean decisions. The Apache 2.0-licensed model has 144.3 million parameters and can run on a laptop CPU or in a browser through a separate WebGPU export.
Julia-1 is the first model in the Julia family and the first public result from Supersonic Labs’ training system. It builds on JHU CLSP’s mmBERT-small, a multilingual ModernBERT encoder that converts text into numerical representations. Supersonic added a classification layer that scores candidate answers and trained the combined model on decision-format examples. The release includes model weights and inference code, while the training pipeline remains private.
Labels arrive with each request
At inference time, Julia-1 receives a state, a question, and candidate options, then scores those options in their supplied order. Candidate meanings form part of each request, allowing an application to change labels without retraining the model or deploying a new output head.
| Mode | Purpose | Output |
|---|---|---|
choice |
Select among two to 20 labeled options | Winning option ID and probabilities |
score |
Evaluate an ordered rubric | Rubric index |
noul |
Make a Boolean decision, with optional descriptions | true or false |
The Python package loads an inference engine for CPU execution. This example routes a duplicate-charge report to one of three teams:
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
