SupersonicLabs/Julia-1

SupersonicLabs releases Julia 1, a 144.3M parameter model for context-grounded decision tasks. The repository documents H200 BF16 inference and provides benchmark results across classification and routing pilots. Performance varies significantly by dataset, with strong results on AG News but notable gaps in Banking77. The release serves as a test of the publisher's training system.

SupersonicLabs presents Julia 1 as the initial entry in a new model family, containing 144.3M parameters. The architecture accepts a specific state, a question, and a list of answer options to produce a single selected decision. This design supports classification, routing, and Boolean tasks without requiring separate output structures for each workflow type. It relies on the mmBERT-small encoder foundation for multilingual processing capabilities. Operators configure the model by providing explicit candidate answers alongside the input context. The system compares the supplied options against the grounded state to determine the most appropriate choice. Users can leverage a unified interface for scoring, ranking, or simple decision making. The runtime supports processing up to 8,192 combined tokens, though historical benchmarks utilized fewer characters. Performance metrics vary significantly across different datasets. The publisher reports strong results on AG News with 94 correct selections out of 100 pilot cases. However, the Banking77 pilot showed a notable gap, achieving only 64 correct answers compared to a reference standard of 87. The team notes that the model cannot supply missing facts or perform complex algebraic reasoning. It is uncertain how the system will handle ambiguous wording or unfamiliar domains in live environments.

README

SupersonicLabs/Julia-1 View on Hugging Face

Loading the README from Hugging Face…