Supersonic Labs Publishing the lightweight decision model Julia-1 and stating that it can run on Android tablets

AT A GLANCE

Supersonic Labs released the lightweight decision model Julia-1, stating that it can run on Android tablets and assist with Agent decisions.

Article

parameters is enough to make Agent decisions, Julia-1 even runs on Android tablets

Beating AI News reports that AI has Supersonic Labs released a lightweight decision model with only 144300000 parameters, Julia-1. It does not generate long-form text; instead, it receives context, a question, and 2 to 20 candidate answers and directly performs classification, scoring, and yes-or-no judgments. It can Agent select tools, perform routing, or decide the next action.

Julia-1 was developed based on the multilingual encoder model mmBERT-small. Its model weights are approximately 550MB. It is designed for low hardware requirements. Official tests have been completed on Apple M4, Intel i5-1235U, and a Samsung Android tablet. The median latency for a single decision on the M4 was approximately 33ms; on the Samsung tablet, using only CPU at runtime, it was approximately 203ms, with peak process memory of approximately 393MB.

In the official 4-group Jev comparison tests, Julia-1 won 3 groups. Typed Decisions scored 73.15%, while Jev's reference score was 72.70%；AG; News scored 94%versus 91%; and sentiment classification scored 86%versus 48%. However, on Banking77, which contains 72 categories, Julia-1 scored only 64%, clearly below Jev's 87%. The official explanation is that when there are too many candidate categories and they are too similar to one another, the current grouped filtering may discard the correct answer prematurely.

Julia-1 cloud GPU training and experiments cost only approximately 104 USD. The model weights, runtime code, complete evaluations, and source records have all been released under the Apache 2.0 license. The official team is also developing API and plans to charge token 0.025 USD per million input token; because the model only makes judgments and does not generate long-form text, output https://m.theblockbeats.info/flash/369158

Key points

The report stated that Supersonic Labs released a lightweight decision model with 144300000 parameters, Julia-1。
Julia-1 receives context, a question, and 2 to 20 candidate answers for classification, scoring, and yes-or-no judgments, without generating long-form text.
Julia-1 was developed based on the multilingual encoder model mmBERT-small, with model weights of approximately 550MB。
The report said that official tests had been completed on Apple M4, Intel i5-1235U, and a Samsung Android tablet.
Official test data showed that the median latency for a single decision on the M4 was approximately 33ms; on the Samsung tablet, it used only CPU at runtime and had approximately 203ms latency, with peak process memory of approximately 393MB。
In the official 4-group Jev comparison tests, Julia-1 won 3 groups: Typed Decisions scored 73.15%versus 72.70%，AG News scored 94%versus 91%, and sentiment classification scored 86%versus 48%。
On the Banking77 test, which contains 72 categories, Julia-1 scored 64%, below Jev's 87%。
The official explanation is that when there are too many candidate categories and they are too similar, grouped filtering may discard the correct answer prematurely.
The report said that Julia-1's cloud GPU training and experiments cost approximately 104 USD。
The model weights, runtime code, complete evaluations, and source records have been released under the Apache 2.0 license.
The official team is developing API and plans to charge token 0.025 USD per million input token; output

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

Julia-1 is not a chat model; its primary task is to make decisions among a limited set of candidates. For example,Agent when faced with multiple tools or next actions, it can be tasked with selecting, ranking, or deciding whether to execute them. The official tests cited in the report show that it can run on some personal-computer and Android-tablet hardware, but it performs less well when there are many categories that are similar to one another.

Why it matters to readers

If the officially reported speed, memory usage, and accuracy can be independently reproduced, developers may be able to complete Agent tool selection and task routing on local devices with a smaller model, reducing reliance on high-performance hardware. Open weights, code, and the Apache 2.0 license would also facilitate testing and integration.

Beginners can understand Julia-1 as a small model responsible for “making decisions among several options,” rather than a chatbot responsible for writing long articles. It may be suitable for learning local Agent routing, but its accuracy limitations on multi-category tasks are worth testing first.

Risks and unknowns

Not provided in the materials Supersonic Labs The original announcement, code repository, or model-weight links were not provided.
Performance, memory, accuracy, and approximate 104 USD costs are mainly based on officially reported and relayed data; no independent retest results were provided.

Supersonic Labs Publishing the lightweight decision model Julia-1 and stating that it can run on Android tablets

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center