Supersonic Labs Launches Julia-1, a 144M-parameter Lightweight Decision Model

ME AI News, Dongcha Beating AI Bulletin: AI lab Supersonic Labs has released Julia-1, a lightweight decision model with only 144.3 million parameters. It does not generate long text; instead, it accepts context, questions, and 2 to 20 candidate answers to directly perform classification, scoring, and yes/no judgments—ideal for agents selecting tools, routing decisions, or determining next steps. Julia-1 is built on the multilingual encoding model mmBERT-small, with model weights totaling approximately 550 MB and designed for low hardware requirements. Testing has been completed on Apple M4, Intel i5-1235U, and Samsung Android tablets. On the M4, the median latency per decision is about 33 ms; on the Samsung tablet running solely on CPU, it is approximately 203 ms, with peak process memory usage around 393 MB. In four official Jev benchmark tests, Julia-1 outperformed Jev in three: Typed Decisions at 73.15% vs. Jev’s 72.70%; AG News at 94% vs. 91%; sentiment classification at 86% vs. 48%. However, on Banking77—with 72 categories—Julia-1 scored only 64%, significantly lower than Jev’s 87%. The team notes that when candidate categories are too numerous and too similar, the current filtering approach may prematurely discard correct answers. The total cost for cloud GPU training and experiments for Julia-1 was only about $104. The model weights, running code, full evaluation metrics, and source records are all open-sourced under the Apache 2.0 license. The team is also developing an API, planned to charge $0.025 per million input tokens; since the model only makes judgments and does not generate long text, output tokens are free. (Source: BlockBeats)