Seven hundred and ninety-nine repositories carry the system-one tag on Hugging Face, and most of them are not for you. The way to cut the pile is a short sequence: the interface decides whether it drops into your code, the size decides where it runs, the license decides whether it can ship, and only then do the benchmark claims matter.
Start with the interface
If your code already speaks Jev's POST /v1/systemone, the shortest path is a model that serves the same contract. Interfaze Lev does, from a 4B LoRA adapter on Qwen3.5-4B. OpenDecider Nano does, from a 400M encoder. Drex DLM does, from an 8B diffusion backbone with its own hosted twin. bev-decider, Darwin-27B-ZTC, and Strands Decider 2B all ship a serve command that exposes it.
Everything else has its own shape: AutoTrust's students expose POST /v1/decide beside an OpenAI-compatible chat path, Liquid d1-3B uses a Python system_one() method with no HTTP decision route, and Julia 1 is a package, not a server. Converting is usually a thin adapter, but it is work your SDK will not do for you.
Match the model to the job
| Model | Size | Where it fits |
|---|---|---|
| Cua-S1 Forms | 706K | form-field choices, browser-side |
| Julia 1 | 144M | 2–20 options, low latency, multilingual |
| OpenDecider Nano | 400M | browser or 4 GB GPU, Jev-shaped API |
| Laya | 421M | English or 100+ languages, guardrails |
| Strands Decider 2B | 2B | agent routing, tool selection |
| Drex DLM | 8B | non-commercial, Mac-friendly |
| JEV-9B / JEV-27B | 9B / 27B | distillation fidelity to hosted Jev |
Option counts discriminate too: Julia accepts 2–20 options natively, d1's score questions take 2–10 levels, AutoTrust's decision head reads up to 256 options, and Laya's fixed head budget starts dropping accuracy past roughly 20. If your label space is 77 categories, that limit decides more than the accuracy table does.
Read the license before the benchmark
Most of the family is Apache-2.0. The exceptions matter: Drex's weights are CC BY-NC 4.0, bev-decider is CC BY-NC because of its training data, d1-3B is under Liquid's lfm1.0 terms, and Security-One 27B is gated behind a conditions gate. Jev-Style v3 is Apache-2.0 but notes some training rows carry restrictive or unclear terms — read the card's data section, not just the license line.
Check the claims
Then read provenance. Mapika's readout scored 57.33 on Decision Index v0.2.1, second of 70 and just under Jev 1.13.0's 57.91 — and the number was measured by the leaderboard rather than the project. JEV-27B-VL tops the community vision board at 69.82, above a stock 397B reference. Both are real signals, and both are contestable in ways the cards disclose.
Three questions sort the claims quickly. Was the comparison fine-tuned against fine-tuned, or fine-tuned against zero-shot? Were the Jev reference numbers published by TypeSafe, or rerun by the author? And is the headline accuracy or calibration — the Typed Decision Leaderboard ties three systems within a bootstrap interval, which is a more honest result than a clean ranking would be.
Pick the interface that fits, verify the license, reproduce one claim on your own split, and keep the hosted API in your back pocket as the reference. The full tour is in Open weights on Hugging Face, and the calibration pass is in Fine-tuning and calibration.