Supersonic Labs releases a 144M-parameter classifier built for CPU inference
Julia 1 is open under Apache 2.0, but Supersonic's own small Banking77 pilot scored below its comparison benchmark.
By Ryan Merket · Published
Primary source: X
Why it matters
Julia 1 is a small, open model aimed at local classification and routing, where CPU deployment could avoid a hosted inference call. Its release includes useful implementation detail, but its own Banking77 pilot shows why teams will need to test the model against their actual labels and workloads.
Supersonic Labs released Julia 1 on September 26th, a 144.3-million-parameter model designed to classify and route text on CPUs. The Brazil-based lab introduced it in a post on X and published the weights and code on Hugging Face under an Apache 2.0 license.
Julia 1 takes context, a question and a set of candidate answers, then selects an option. Supersonic says it can handle classification, routing, ordered scores and yes-or-no decisions. It is not a text-generation model: the system chooses among answers supplied by the developer.
The launch fits Supersonic Labs' stated focus on running practical AI locally, on hardware people already own. The model builds on the multilingual mmBERT-small encoder rather than training a foundation model from scratch. Supersonic says cloud GPU costs for Julia 1's training and experiments totaled about R$540, or $104.08.
The benchmark results are mixed. In a 100-example Banking77 pilot with 72 labels narrowed through a shortlist, Julia 1 got 64 examples right; the comparison figure Supersonic supplied was 87%. It scored 94 of 100 on a four-label news pilot and 86 of 100 on a six-label emotion pilot. Those small pilots are company-reported results, not evidence of performance across production workloads.
The model's 550.5 MiB weight file can run on a CPU, and Supersonic also reports tests on an Apple M4, an Intel Core i5-1235U laptop and a Samsung tablet. The lab reports five decisions per second on the tablet, using CPU execution after its available accelerator path fell back to CPU. That is a deployment demonstration, not a guarantee that Julia 1 will run efficiently on every device.
Supersonic says an API is in development, with planned pricing of $0.025 per million input tokens and no charge for output tokens. For now, the open release gives developers a model to test locally; the company's own results point to long lists of similar choices as a meaningful weakness.