Supersonic labs launches julia 1: a compact, cpu-driven decision model with 144.3M parameters

Supersonic Labs, a Brazilian AI research lab, has unveiled Julia 1, a decision-making model with 144.3 million parameters that operates on standard CPUs. Unlike chatbots, Julia 1 processes context, a question, and multiple answer choices, returning probabilities for each option. The model is open-source, available on Hugging Face under Apache 2.0, and supports deployment on Python 3.11+ environments or BF16-capable GPUs. An ONNX version runs in browsers via WebGPU, and a hosted API is in development. Julia 1 is based on the mmBERT-small encoder, trained on over 1,800 languages, with a decision-making head added by Supersonic Labs. It handles three decision types through a single API, providing softmax probabilities in the order of input options. Benchmarks show it achieves 72.55% accuracy on Typed Decisions and 60/100 on Banking77, with some abstentions. Performance varies across devices, with the Apple M4 achieving a 33.15 ms median response time. The model is not a fine-tuned version of Qwen and is designed for classification tasks, though it cannot handle missing facts or complex calculations. Supersonic Labs advises users to validate results with human oversight for critical decisions. The lab also announced plans for Julia 2, featuring its own foundation architecture.