Supersonic Labs Releases Julia 1 CPU-Runnable Decision Model

Supersonic Labs, an AI laboratory based in Brazil, has released Julia 1, a compact decision model designed for specific tasks rather than general conversational AI. This model accepts context, a question, and between two to twenty candidate answers, returning a selected answer along with a probability for each option. Julia 1 features 144.3 million parameters and is engineered to operate on a standard CPU, making it accessible for local deployment. The model's weights are publicly available on Hugging Face under the Apache 2.0 license, supporting execution with Python 3.11 or later on a CPU or a BF16-capable GPU. Furthermore, an ONNX build enables browser-based execution via WebGPU. While a hosted API has been announced, it is not yet publicly accessible.

Julia 1 is structured to handle three distinct decision-making types through a unified API: 'choice' for selecting one label from a list of 2 to 20 options (applicable to classification and routing tasks), 'score' for returning an index on an ordered scale such as low, medium, or high, and 'noul' for determining the probability of a true yes-or-no statement. The model returns results in the order specified by the caller, including full softmax probabilities, and preserves caller-defined identifiers like billing information. Crucially, Julia 1 does not generate text.

The architecture of Julia 1 is based on JHU CLSP’s mmBERT-small, a multilingual ModernBERT encoder with 140 million parameters trained on over 1,800 languages. Supersonic Labs retained the encoder and tokenizer from this base model, appended a decision-specific head, and conducted training using examples formatted for decision-making tasks. The lab has clarified that Julia 1 is not a fine-tuned version of a Qwen model. The runtime environment supports up to 8,192 combined tokens, though published benchmarks utilized a 1,024-token limit. The total expenditure for cloud GPU resources for training and experimentation was approximately R$540, equivalent to US$104.08. The FP32 weights for the model occupy 550.5 MiB of storage. The private training pipeline used by Supersonic Labs has not been released. The lab is currently developing Julia 2, which will feature its own proprietary foundation architecture.

Benchmark evaluations conducted on September 24, 2026, using H200 BF16 with strict encoding, positioned Julia 1 against TypeSafe’s Jev, using reference values from the Jev benchmark protocol. In the 'Typed Decisions' benchmark, Julia 1 achieved 73.15% accuracy (1,463 out of 2,000 instances), compared to the reference value of 72.70%. For the AG News dataset with 4 labels, Julia 1 scored 94%, surpassing the reference of 91%. In the DAIR Emotion benchmark, which involves 6 labels, Julia 1 reached 86% accuracy, a significant improvement over the reference score of 48%.