Supersonic Labs releases Julia-1, a 144M decision model that runs on CPU
On 26 September 2026, Brazilian lab Supersonic Labs announced Julia-1 on X (@supersonicai, 16:00 BRT): “our first classification model that runs on almost anything”. The weights had been on Hugging Face since 23 September.
Julia-1 is a 144.3M-parameter fine-tune of mmBERT-small, a multilingual ModernBERT encoder, with a decision head. It takes a state, a question and 2 to 20 options and returns one pick with a probability per option, through a single typed API for choice, score and noul. Apache-2.0, FP32 weights of 550.5 MiB, Python runtime on CPU, plus an ONNX/WebGPU build. The lab stresses it is not a Qwen fine-tune.
What the lab reports (UnverifiedClaim): 73.15% on the 2,000 typed decisions of LocalLLaMA/typed-decisions against a supplied Jev reference of 72.70%, strong 100-example pilots on AG News (94) and DAIR Emotion (86), and a clear miss on a 72-label Banking77 shortlist (64 vs 87). Median latency of 33 ms per decision on an Apple M4 and 295 ms on an Intel i5 laptop. About R$540 (US$104) of cloud GPU spend in total. A hosted API at $0.025 per million input tokens is planned, not live. Julia 2, on the lab’s own architecture, is in development.
How to read it: the Jev column is a reference value from the benchmark protocol, not a new Jev run. DecisionEval independently measured Jev 1.13.0 at 0.740 (CI 0.721–0.759) on the same split, so on typed decisions Julia-1 is level with Jev rather than ahead. The pilots are 100 examples each, and the checkpoint’s provenance file lists the same task families among its validation sources. No calibration step ships (calibration: null). Launch-day traction: 161 likes on the X post a few hours in, 13 likes and 0 downloads on HF.
What makes the release worth a look is the paperwork: a pinned protocol, per-benchmark counts, a provenance file and published weaknesses, on a model small enough for anyone to rerun on a laptop.
Catalog entry: /models/julia-1. Nearby peers: Lumma-fev, Tiny-Jev, Laya.