Search the Hub for system-one and you get 799 model repositories. That number is misleading in both directions: most are variants of a dozen ideas, and the interesting ones are not the biggest. Here is what is actually on Hugging Face, and what a download buys you compared with the hosted API.
What the Hub actually has
The families are easy to separate once you look. Laya carries three checkpoints in one repository: a 421M English encoder, a 322M multilingual build, and a fine-tuned typed-decisions model. Julia 1 is a 144.3M encoder with a WebGPU export next door. AutoTrust's JEV-27B, JEV-27B-VL, JEV-9B, and GEV-26B-Decide are distilled students on frozen Qwen and Gemma backbones. Liquid AI's d1-3B and Mapika's decider readout round out the mid-size tier.
Beside the models sit 21 datasets with typed-decisions in the name, from the 2,000-decision benchmark to 2.5 million recast decisions, plus two operating leaderboards: the Typed Decision Leaderboard and the S1MB mosaic board.
Three shapes of open decision model
The first shape is encoder-first: a bidirectional backbone does the language work and a decision head reads options from its hidden states. Laya, Julia 1, and OpenDecider Nano all fit this pattern, and they are the cheapest to run because they answer in one forward pass with no generated tokens.
The second shape is an adapted LLM. AutoTrust's students keep the backbone frozen and train only a rank-16 LoRA plus a 24-slot head — 108.9M of 27B parameters — so System 2 chat still works through the ordinary weights. Mapika goes further: its Gemma-4-31B entry contains no training at all, just a config that reads option letters at the answer slot.
The third shape is a port. Jev-Style v3 ships GGUF builds down to 0.53 GB, and Julia 1's ONNX export runs in a browser through WebGPU. Ports are where most deployment experiments start.
What the model card tells you
Licenses are the first filter. Most of the ecosystem is Apache-2.0, but Drex weights are CC BY-NC, bev-decider is CC BY-NC because of its training data, d1-3B is under Liquid's own lfm1.0 terms, and Security-One is gated behind conditions.
Provenance is the second. Laya's fine-tuned checkpoint scores 0.766 on typed decisions, above Jev's published 0.727 — but the card says the Jev figure is third-party published, not a run made there. AutoTrust's KL and accuracy numbers are self-reported, with one independent human-labelled rerun cited alongside. Julia's 73.15% edges a supplied reference table. None of this makes the numbers wrong; it makes them readable.
What open does not include
There are no official Jev weights to download, and no card ships its training pipeline: AutoTrust publishes weights, adapter, and reports; Julia publishes the checkpoint and inference code. Calibration is your job on most of these — Laya arrives over-confident and expects you to refit temperatures, and AutoTrust tells you to apply head bias and temperature client-side.
That is the honest trade. A download gives you a checkpoint, a license, and a runnable interface. Turning it into a dependable threshold is the next article: fine-tuning and calibration. For what people have already built on top of these weights, see Open reimplementations.