# Julia 1

> A compact 144.3M-parameter multilingual decision model from Supersonic Labs that classifies, ranks, and answers yes/no questions by choosing among supplied options, running on CPU.

Julia 1 is a compact decision model from Supersonic Labs, a Brazilian AI research lab, built on the mmBERT-small multilingual encoder. With 144.3 million parameters and a 550.5 MiB checkpoint, it is designed to run on a CPU — including Android tablets — making it accessible without GPU infrastructure. The model is publicly available on Hugging Face under the Apache 2.0 license, and Supersonic Labs states it plans to open API access soon.

## What It Is

Julia 1 is a typed decision model: given a context, a question, and between 2 and 20 possible answers, it selects one and returns scores in the order the options were supplied. This single interface covers classification (routing a support message to a department), ordered scales (rating urgency as low, medium, or high), and yes/no questions. The design goal, as stated by Supersonic Labs, is to study whether one model can handle several kinds of decision tasks as the option set changes.

The model is the first public output of the Julia research line. Supersonic Labs describes Julia 2 as in development, with plans to build a proprietary foundation architecture rather than adapting mmBERT.

## Evaluation Results

Supersonic Labs published benchmark results from a September 24, 2026 evaluation using the Jev protocol:

- **Typed Decisions (2,000 examples):** 73.15% accuracy, 0.45 percentage points above the Jev reference (72.70%)
- **AG News (4 labels, 100 examples):** 94% accuracy vs. 91% reference
- **DAIR Emotion (6 labels, 100 examples):** 86% accuracy vs. 48% reference — a 38-point margin
- **Banking77 (72-label shortlist, 100 examples):** 64% accuracy vs. 87% reference — a concrete failure case with long, similar-category lists

On the MASSIVE multilingual dataset (52 locales, 18 scenarios), Julia 1 answered 110,573 of 154,648 cases correctly (71.50%), with 86.25% on European Portuguese and 86.75% on US English.

## Hardware Performance

Julia 1 has been measured across three device classes:

- **Apple M4 Mac:** 33.15 ms median per decision (single), 51.20 decisions/second in batches of 16; 370.6 MiB RAM
- **Samsung SM-X510 tablet (Android 16, ONNX Runtime):** 5 decisions/second, 203 ms median, 393.1 MB peak RSS; the 550.1 MB weight file was memory-mapped
- **Intel Core i5-1235U:** 294.81 ms median on typed decisions; Banking77 narrowing adds significant latency (3,713 ms median)

The Python 3.11+ runtime accepts up to 1,024 tokens for context, question, and options combined. A Router component can narrow large candidate lists in groups, though the Banking77 results show this narrowing can lose the correct answer.

## Architecture and Training Approach

Julia 1 adapts mmBERT-small — a multilingual encoder from JHU CLSP — by adding decision-scoring components trained to choose among answers supplied with a question. Supersonic Labs states the total cloud GPU spend on Julia 1 training and experiments was approximately R$540 (about US$104), reflecting the lab's constraint of limited funding and compute. The team chose mmBERT-small specifically to avoid the data, infrastructure, and cost of training a multilingual foundation from scratch.

The evaluated configuration uses PyTorch 2.14.0+cpu with strict encoding. An ONNX Runtime path is available for mobile deployment. Checkpoint provenance is tracked via SHA-256 hashes and published alongside the weights on Hugging Face.

## Current Status

Julia 1 weights, Python interface, metrics, and provenance are available on Hugging Face (SupersonicLabs/Julia-1) under Apache 2.0. Supersonic Labs states it is building an API for Julia 1 and plans to open access soon. Julia 2, which the lab describes as targeting a proprietary foundation architecture, is listed as still in development. The project page notes it will be updated with more examples, tests, and information.

## Features
- Multilingual decision model supporting 52+ locales
- Runs on CPU without GPU requirement
- Handles classification, ordered ranking, and yes/no questions with a single interface
- Accepts 2 to 20 possible answers per question
- Up to 1,024 token context window
- ONNX Runtime support for Android/mobile deployment
- Router component for narrowing large candidate lists
- 144.3M parameters, 550.5 MiB checkpoint
- Apache 2.0 licensed weights on Hugging Face
- Checkpoint provenance tracked via SHA-256 hashes

## Integrations
Hugging Face, PyTorch, ONNX Runtime, mmBERT-small tokenizer

## Platforms
MACOS, ANDROID, API, DEVELOPER_SDK, CLI

## Pricing
Open Source

## Version
Julia 1 (evaluated 2026-09-24)

## Links
- Website: https://supersoniclabs.ia.br/julia-1/
- Repository: https://github.com/AbdelStark/jev-benchmarks
- EveryDev.ai: https://www.everydev.ai/tools/julia-1