ME AI message: observe the Beating AI quick news—AI Laboratory Supersonic Labs has released a lightweight decision model, Julia-1, with only 144.3 million parameters. It does not generate long text; instead, it takes the context, a question, and 2 to 20 candidate answers to directly perform classification, scoring, and true/false judgments. It can help Agents choose tools, route requests, or decide the next action. Julia-1 is built on the multilingual encoding model mmBERT-small. The model weights are about 550MB, with the main focus on low hardware requirements. The official team has already tested it on Apple M4, Intel i5-1235U, and Samsung Android tablets. On M4, the median single-decision latency is about 33ms; on a Samsung tablet running only on the CPU, it’s about 203ms, with peak process memory around 393MB. In four officially announced Jev comparison tests, Julia-1 won three of the four groups. Typed Decisions: 73.15%, Jev reference: 72.70%; AG News: 94% vs 91%; sentiment classification: 86% vs 48%. However, on Banking77, which includes 72 categories, Julia-1 scores 64%, significantly lower than Jev’s 87%. The official says that when candidate categories are too many and too similar to each other, the current grouping-based filtering may discard the correct answer too early. The total cloud GPU training and experimentation cost for Julia-1 is only about $104. The model weights, runtime code, complete evaluation results, and source tracking are all open, released under the Apache 2.0 license. The official 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 not charged. (Source: ME)