The Pulse

Supersonic Labs Releases Julia 1, a CPU-Ready Decision Model

Supersonic Labs released Julia 1, a 144.3-million-parameter model that runs on a CPU and chooses among supplied answers. It beat Jev reference scores on three of four reported tests but fell short on a 72-label banking pilot.

AI.info Team · Written with AI · fact-checked · 2026-09-27

Supersonic Labs’ new Julia 1 model beat Jev reference scores on three of four reported tests, the lab says. But on Banking77, a pilot asking it to sort among 72 banking categories, Julia scored 64% against the reference 87%—a sharp limit for a model built to make choices among supplied options.

The Brazilian lab released Julia 1 on September 26 as a 144.3-million-parameter model designed for classification, routing, ordered scoring and yes-or-no decisions, rather than text generation. Its results are promising on several small tests, but the lab’s own evaluations also show where the model struggles: long lists of similar labels and tasks that depend on knowledge not included in the input.

Julia 1 makes choices, not conversation

Julia takes a context, a question and between two and 20 possible answers, then returns a selected option and scores. Supersonic Labs built it on JHU CLSP’s mmBERT-small multilingual encoder, adding a decision head and training it on decision-format examples. The company publishes the model weights and inference code, but not its training pipeline.

“Julia 1 opens our research into compact decision models.”

Supersonic Labs

The weights are available under the Apache 2.0 license. The model runs on a CPU, according to Supersonic’s release page; the repository lists a 550.5 MiB checkpoint and instructions for local use with Python 3.11 or later. A separate ONNX export supports browser inference through WebGPU.

Three benchmark wins, one clear shortfall

In an evaluation dated September 24, Julia scored 73.15% on 2,000 typed decisions, slightly above the supplied Jev reference of 72.70%. It scored 94 out of 100 on an AG News pilot and 86 out of 100 on a DAIR Emotion pilot, above the respective references of 91% and 48%. Those comparison figures come from the Jev protocol; Supersonic did not run a fresh Jev comparison alongside Julia.

The Banking77 pilot exposes the difficulty of expanding the option set. Julia reached 64 out of 100, versus the supplied 87% reference, using a shortlist process that narrows 72 categories before a final decision. Each classification pilot covered 100 examples, so the scores offer a limited test rather than a broad measure of performance across real deployments.

Julia also scored 71.50% macro accuracy on MASSIVE scenario classification across 52 locales, choosing among 18 scenarios, Supersonic reports. The lab cautions that the test covers scenario selection, not intent classification or slot filling. Its model card says results do not establish accuracy in new domains or high-stakes settings.

CPU measurements depend on the task

Median latency was 33.15 milliseconds per decision on an Apple M4, the lab reports, using requests with 100 context words and four options. On an Intel Core i5-1235U, its reported median was 107.83 milliseconds for AG News decisions. Banking77 took 3,713.54 milliseconds per decision in the same set of measurements, reflecting the extra work of narrowing a 72-label list.

Those figures come from different tasks and devices, not a single controlled speed comparison. The company’s page also reports a Samsung tablet run through ONNX Runtime, where inference fell back to the CPU; the median was 203 milliseconds per decision. Julia’s small size makes local execution possible, but it does not remove the accuracy and latency costs of harder choices.

An 8,192-token limit, with an accuracy caveat

A September 27 update to the Hugging Face repository changed Julia’s runtime limit from 1,024 to 8,192 combined tokens for context, question and options. The update describes that limit as matching mmBERT’s native positional capacity, while explicitly warning that accuracy on longer inputs has not been established. The benchmark results reported above used the earlier 1,024-token limit.

Supersonic plans a hosted API, priced at $0.025 per million input tokens, but the service is not yet open. For now, developers can download the Apache-licensed weights and run the model themselves; the company has not published the training pipeline.