Five open decision models from Hugging Face's trending list join the catalog
Five decision models were on Hugging Face’s trending list for text classification on 30 September 2026, and none of them was in the catalog yet. All five return typed answers (Noul, Choice, Score) with probabilities from trained weights, so all five now have catalog pages. They differ a lot in how much of what they claim has been checked by anyone else.
Where we cite the Decision Index, the numbers come from the official Decision Index Space (data snapshot 28 Sep 2026, 71 systems, Jev 1.13 on top at 57.91). Everything else is the authors’ own measurement (UnverifiedClaim).
| Model | Who | What it is | Decision Index 0.2.1 | Quadrant (M / C / A) |
|---|---|---|---|---|
| Surogate Rune 26B-A4B v3 | Invergent | Full fine-tune of Gemma 4 26B-A4B, text + images | 57.44 (#2) | 5.8 / 7.9 / 11 |
| AutoJev-27B | denis-pplx | Full-weight SFT of Qwen3.8-27B, /v1/systemone server |
56.40 (#4) | 5.1 / 7.2 / 11 |
| Mica v0.1 4B | sky7350 / akivet | Merged LoRA on Qwen3.5-4B, English + Korean, GGUF | not listed | 5.4 / 7.3 / 8 |
| NeoHorse-Jev-4B | TokenRhythm | Kev-style pointer head trained on NeoHorse-1-4B | 36.75 (#30) | 6.1 / 5.7 / 9 |
| AutoTrust JEV | AutoTrust AI | LoRA + decision head on a frozen Qwen, distilled from Jev | not listed | 6.1 / 7.5 / 9 |
The two near the top. Rune is 0.47 points behind Jev on the index, in a run the maintainers reproduced, and ahead of it on language, retrieval and arts. It is overconfident at its default temperature (index ECE 0.12). AutoJev-27B is 1.5 points behind Jev, with the best measured calibration of the five (index ECE 0.018).
The self-report gap. NeoHorse-Jev-4B’s card leads with a 77.70 six-group average and “highest among open models”; the index puts it at 36.75, just above Kev 4B. Mica publishes per-item JevBench outputs you can re-check, but it is not on the index yet.
The distillation. AutoTrust’s JEV models were trained on a third-party corpus whose main stream is labelled with Jev 1.13 outputs obtained through OpenRouter. Their headline numbers measure how closely they copy Jev, not accuracy against ground truth. The name is also easy to confuse: “JEV” here is not TypeSafe’s product, and AutoTrust says it has no affiliation. The page carries both caveats at the top. We have not verified whether training on Jev outputs complies with TypeSafe’s terms.