Supersonic Labs’ Julia 1 Fits Typed Decisions Into a 144M CPU Model

Supersonic Labs released Julia 1 as a 144.3-million-parameter open decision model that runs on ordinary CPUs and chooses among answers supplied with a question. The company’s launch page and Hugging Face model card describe a single interface for classification, routing, ordered scores, and yes-or-no (noul) decisions: pass context, a question, and two to twenty options; get back scores in the caller’s option order and one selected answer. Weights and the Python runtime are Apache License 2.0. The foundation is JHU CLSP’s mmBERT-small multilingual encoder; Julia adds a decision head rather than shipping as a chat or masked-language model drop-in.

Measured results are dated and reproducible enough to cite carefully. On a September 24, 2026 evaluation with H200 BF16 inference and strict encoding, Julia 1 answered 1,463 of 2,000 typed decisions correctly (73.15%), 0.45 percentage points above the protocol’s supplied Jev reference of 72.70%. Split by type, Choice was 428/600, Noul 484/600, and Score 551/800. One-hundred-example pilots scored 94/100 on AG News (four labels) and 86/100 on DAIR Emotion (six labels), while a Banking77 pilot that narrows seventy-two labels through a shortlist before the final pick reached only 64/100 against an 87% reference—Supersonic calls long, similar category lists an explicit weakness. MASSIVE scenario classification across fifty-two locales landed at 110,573 of 154,648 correct (71.50%), with Portuguese (pt-PT) and English (en-US) both above 86% on the eighteen-scenario task.

A separate CPU package recorded on September 25, 2026 used PyTorch 2.14.0+cpu, strict encoding, and a 1,024-token limit. Typed decisions were 1,451/2,000 (72.55%); AG News and Emotion pilots matched the GPU pilots at 94 and 86; Banking77 fell to 60/100 with three abstentions counted among the hundred cases. Latency tables on the company page show an Apple M4 at about 33 ms median per single decision (roughly 28 decisions/s) and about 51 decisions/s when batched sixteen at a time; an Intel Core i5-1235U run of the full typed suite at roughly 295 ms median; and a Samsung SM-X510 tablet ONNX Runtime path at about 203 ms median over forty decisions. Checkpoint provenance is pinned by a SHA-256 prefix (df853bf7fe42…) in the evaluation artifacts. FP32 weights occupy about 550.5 MiB.

The product cut is intentionally narrow. Julia compares the options you provide; it is not presented as a source of missing world knowledge or multi-step arithmetic. Native calls still accept two to twenty options; a hierarchical Router can narrow larger lists in groups, but grouped probabilities cover only finalists and the Banking pilot shows narrowing can drop the right answer. The Hugging Face card documents both a named-question predict(state=..., questions=...) API and a legacy list API, with strict encoding that rejects overflow against configurable token budgets (historical benchmarks used 1,024 combined tokens; runtime supports up to 8,192 with 8k task accuracy not established). A scripts/reproduce_typed.py harness verifies the published weight hash and a pinned typed-decisions Parquet before rewriting predictions.

Resource constraints are part of the disclosure, not a footnote. Supersonic says early GPU cloud spend on training and experiments was about R$540 (US$104.08), choosing mmBERT-small because training a multilingual foundation from scratch was out of budget. Julia 2 is described as future work on a foundation without mmBERT. A hosted API is planned at $0.025 per million input tokens and $0.00 per million output tokens, with access not yet open at launch. ONNX and WebGPU adapters live in a separate Julia-1-ONNX repository for browser and tablet paths.

Relative to larger open decision models and closed Jev-style APIs covered earlier in September, Julia 1’s claim is different: a CPU-first, Apache-2.0 encoder decision head small enough for laptops and tablets, with published failures on long label lists rather than a promise of general routing intelligence. Teams should evaluate on their own option wording before consequential use—the authors say so explicitly.

For Writeble’s products desk, Julia 1 is the compact late-September decision release that makes “runs on a CPU” the headline instead of another multi-billion-parameter distill of a hosted System-1 API.