Live site · by Supersonic Labs

Julia 1

A 144.3M-parameter decision model that edges Jev on typed decisions.

Supersonic Labs' first Julia model, an Apache 2.0 checkpoint built on JHU CLSP's multilingual mmBERT-small. It takes a state, a question and 2 to 20 options and returns one decision, for classification, routing, scores and yes/no. On 2,000 typed decisions it scored 73.15% against a supplied Jev reference of 72.70%, and 94 of 100 on an AG News pilot, but 64 of 100 on a 72-label Banking77 pilot where the reference got 87. It runs on CPU from 550.5 MiB of weights, and an ONNX build runs on WebGPU. The authors say plainly that it compares the answers you give it and cannot supply missing facts.