Julia 1

Julia 1

Takes a text, a question and a list of possible answers, and picks the best one for classification, routing, ordered scores and yes-or-no calls. A 144M model for developers who test routing in PyTorch.

Made by SupersonicLabs on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Julia 1 is a small model with 144M parameters that makes one decision from three inputs. You give it the context, a question and the possible answers. The same interface covers sorting text into labels, routing, ordered scores and yes-or-no calls. The card presents it as the opening release of the Julia family and a first trial of the makers' training setup.

Who it suits: Developers who want to try model-based routing and classification with explicit candidate answers, under the Apache 2.0 license.

Industry | Who uses it | How they use it | Example
Support | Support platform teams | Route a message to the right team | Send a billing complaint to billing
Software | App developers | Label text with a fixed set of categories | Tag a news item as sports or business
Operations | Workflow builders | Ask a yes or no question about a record | Decide whether a ticket is urgent
Research | ML researchers | Test a compact decision model on new label sets | Compare accuracy with and without a shortlist

Skip it if you need the model to supply missing facts or work through long calculations. The card says those abilities are not shown by its tests. Skip it too if you need a final answer without checking it on your own questions first.

Where to use it: It is a PyTorch model that you run in Python.

• PyTorch, with Safetensors weights
• Windows, macOS or Linux, with Python
• A script or service that routes text to a team, queue or label
• A larger label list through the hierarchical Router the card describes

Use case | Who | What they get back
Labeling | Developers | The best label from your list
Routing | Platform teams | The team or queue that fits the message
Yes or no checks | Workflow builders | A yes or no decision
Ordered scores | Analysts | A level on an ordered scale

How to set it up: You need Python and PyTorch.

• Python with pip
• PyTorch to load the model
• About 577 MB of space for the F32 weights

1. Install the Hugging Face hub tools with pip.
2. Download the model files with the hf command.
3. Follow the card for the code that loads the model and builds a request.

pip install -U huggingface_hub
hf download SupersonicLabs/Julia-1

How to use it: Build a request with the context, one question and the candidate answers. The request can be a choice among options, an ordered score, or a Boolean noul decision. One native call accepts 2 to 20 options. The card says to use strict encoding. The card text shown here ends before the code, so follow the card for the exact calls.

Tool | What it does
choice | Picks one answer from 2 to 20 options
score | Gives a level on an ordered scale
noul | Answers a Boolean question
Router | Narrows a long choice list by groups and reranks the survivors

Example: The context is a customer message that says a card was charged twice. The question asks which team should handle it, and the options are billing, technical support and sales. The model picks the option that fits best.

For newcomers and experts: If you are new to this, think of the model as a multiple-choice test taker. It reads the context and compares the answers you offer, and it does not invent new ones. Test it on a handful of your own cases first.

If you know the field, the scores were taken on 2026-09-24 on an H200 GPU in BF16, with strict encoding. Pilots use 100 examples per dataset. The Banking77 pilot runs 72 labels through a top-16 shortlist rather than one native call, and abstentions count as wrong. The Jev numbers are supplied references, not a new run. A grouped Router result is not a global probability distribution. A section on the path from mmBERT-small to Julia 1 is cut from the text shown here.

Write the context
Ask one question
List the candidate answers
Run the model with strict encoding
Read the chosen answer

Good to know: The card reports 73.15 percent on its typed decisions, 94.00 percent on an AG News pilot and 86.00 percent on a DAIR Emotion pilot. The Banking77 pilot reached 64.00 percent, below the supplied reference of 87.00 percent. The card calls the pilots encouraging signals and not guarantees.

Title: Julia 1 accuracy by benchmark
AG News pilot: 94.00%
DAIR Emotion pilot: 86.00%
Typed decisions: 73.15%
MASSIVE, all 52 locales: 71.50%
Banking77 pilot: 64.00%

• License: Apache 2.0
• Cost to run: none named, apart from your own hardware
• Data it reaches: the context and options you give it and the model files on your computer, and nothing else the card names
• Not for: supplying missing facts, algebra or long calculations, and consequential actions you have not tested on your own workflow

On Hugging Face, 13 November 2026

• Task: text classification
• Library: pytorch
• Parameters: 144M
• Weights: about 577 MB in F32; about 72 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 2,162
• Likes: 503
• License: apache-2.0
• Quantized versions: 8
• Fine-tunes: 8
• Tags: pytorch, safetensors, decision-model, text-classification, multilingual, routing

Links

• Model: https://huggingface.co/SupersonicLabs/Julia-1
• Files: https://huggingface.co/SupersonicLabs/Julia-1/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download SupersonicLabs/Julia-1

Then follow the usage in its model card. Nothing in this kit runs by itself.