Supersonic Labs' Julia-1 is a small open decision model that runs on a CPU or in the browser

Supersonic Labs has released Julia-1, a new classification model. The company calls it its first classification model that can run on almost anything. The model is small enough to run on your own computer or server, and its code and model files are public on Hugging Face under the Apache 2.0 license.

What Julia-1 is

Julia-1 is a decision model. You give it a state, a question and between 2 and 20 possible answers, and it picks one and returns a probability for each option. For example, given the message "I was charged twice for the same order" and the question "which team should handle this request", it picks billing over shipping and account access.

The same interface handles three kinds of questions: picking from options, scoring on an ordered scale, and yes-or-no decisions. Several questions can be scored on the same text at once. The model is built on Johns Hopkins University's multilingual mmBERT-small model and trained to make decisions.

What it is for

Julia-1 is mainly designed for routing: sending customer messages to the right team, labeling the topic or emotion of a text, or deciding which tool or model a request should go to. According to the company, it looks at the meaning of the text and the options instead of relying on keyword rules, and it can work with new option lists without extra training for each new task.

Where it runs

Julia-1 runs on an ordinary CPU with no GPU needed. A separate ONNX version lets it run directly in the browser through WebGPU, so text can be classified without leaving the user's computer. The model page includes a step-by-step Python setup guide and sample code.

The results the company shared

According to the company's own measurements, Julia-1 comes out ahead of the reference results given for TypeSafe's decision model Jev on news-topic and emotion classification and on a decision benchmark. The model was also tested on scenario classification in 52 languages, including Turkish. The company itself notes that the model trails Jev on a test that sorts banking requests into many categories. Some of these tests are pilots run on one hundred examples.

Supersonic Labs also lists Julia-1's limits clearly on the model page. The model only compares the options it is given, and it should not be expected to fill in missing facts, solve equations or carry long calculations. Ambiguous wording, unfamiliar domains and long option lists can lead to mistakes. The company recommends testing the model with your own questions and options before using it for important decisions. The training pipeline has not been released. To see how it performs on other tasks, it is worth following evaluations from independent researchers as well.

Sources: Hugging Face: Julia-1, Hugging Face: Julia-1-ONNX.