Supersonic Labs Releases Julia 1, a 144.3M-Parameter Decision Model That Runs on CPU

Julia 1 handles three decision types through one API: choice, which picks one label from 2 to 20 described options for classification or routing; score, which returns the expected index on an ordered rubric such as low, medium, or high; and "noul," which returns the probability that a yes-or-no statement is true. Results come back in the caller's option order with full softmax probabilities, and caller IDs such as billing are returned unchanged. The model does not generate text.

The model starts from JHU CLSP's mmBERT-small, a 140-million-parameter multilingual ModernBERT encoder trained on more than 1,800 languages. Supersonic Labs kept the encoder and tokenizer, added a decision head, and trained on decision-format examples. The lab states that Julia 1 is not a fine-tuned Qwen model. The runtime supports 8,192 combined tokens, though published benchmarks used a 1,024-token limit. Total cloud GPU spending for training and experiments was about R$540, or US$104.08. The FP32 weights occupy 550.5 MiB, and the private training pipeline has not been released. Julia 2, built on the lab's own foundation architecture, is in development.

A September 24, 2026 evaluation ran on H200 BF16 with strict encoding. The comparison baseline is TypeSafe's Jev, using reference values from the Jev benchmark protocol rather than a new Jev run. On Typed Decisions, Julia 1 scored 73.15 percent (1,463 of 2,000) against a 72.70 percent reference. On AG News with four labels, it scored 94 out of 100 against a 91 percent reference. On DAIR Emotion with six labels, it scored 86 out of 100 against 48 percent. On Banking77 with 72 labels, it scored 64 out of 100 against 87 percent, which the report describes as a clear failure. On MASSIVE with 18 scenarios, it reached 71.50 percent macro accuracy across 52 locales, including 86.25 percent for pt-PT and 86.75 percent for en-US. The classification pilots used only 100 examples each. A September 25 CPU run reproduced most numbers, with 72.55 percent on Typed Decisions and 60 out of 100 on Banking77 with three abstentions.

The lab published per-device latency measurements. On an Apple M4, one decision per call took a 33.15 ms median. On a Samsung SM-X510 tablet via ONNX Runtime, the median was 203 ms with 393.1 MB peak RSS. On an Intel Core i5-1235U, AG News decisions took a 107.83 ms median, while Banking77 took 3,713.54 ms because it narrows 72 labels first. On X, @supersonicai claims Julia 1 classifies five times faster than Jev on an i5 laptop; the report cautions that the Jev pilot measured Jev as a hosted service called from France, so the latencies are not like-for-like.

Julia 1 compares the answers supplied and cannot be relied on for missing facts, algebra, or multi-step calculation. The Router can drop the correct label during narrowing. It is not a drop-in Transformers pipeline, and no Hugging Face inference provider serves it. Supersonic Labs advises evaluating on your own questions and keeping humans in the loop for consequential decisions. A hosted API is planned at $0.025 per million input tokens, but it is not yet open.