144M parameters are enough for Agent decision-making, and Julia-1 can even run on Android tablets.

Beating AI News Flash: 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 takes context, a question, and 2 to 20 candidate answers, directly performing classification, scoring, and true/false judgment, and can help Agents choose tools, perform routing, or decide the next action.

Julia-1 is built on the multilingual encoder model mmBERT-small, with model weights of about 550MB, focusing on low hardware requirements. The official team has already completed tests on Apple M4, Intel i5-1235U, and a Samsung Android tablet. On M4, the median latency for a single decision is about 33ms; on the Samsung tablet running only on CPU, it is about 203ms, with peak process memory of about 393MB.

In the 4 official Jev comparison tests, Julia-1 won 3. Typed Decisions was 73.15%, versus Jev's reference score of 72.70%; AG News was 94% versus 91%; sentiment classification was 86% versus 48%. However, on Banking77, which includes 72 categories, Julia-1 scored only 64%, significantly lower than Jev's 87%. The official team said that when there are too many candidate categories and they are similar to one another, the current grouped filtering may discard the correct answer in advance.

The total cost of Julia-1's cloud GPU training and experiments was only about $104. The model weights, runtime code, full evaluation, and source records have all been open-sourced under the Apache 2.0 license. The official team is also developing an API and plans to charge $0.025 per million input tokens; because the model only makes judgments and does not generate long text, output tokens are not charged.