AutoTrust-JEV-27B v1.1: self-reported 67.31 on the Decision Index public suite, pending verification

Source · Model card: AutoTrust GEV-26B-Decide

publicado (OK do Fabricio 20/10 BRT).

AutoTrust AI published AutoTrust-JEV-27B v1.1 on Hugging Face on 17 October 2026 (04:59 BRT) and announced it on X on 19 October (12:00 BRT). It is a 27B text decision model (about 27.4B parameters in the repo) with Apache-2.0 weights and a decision readout head, from the same company as GEV-26B-Decide.

Serving. The card installs vllm==0.31.0, transformers==5.17.0 and StartLux’s startlux-decide engine at a pinned commit, and turns on its prefix-bidirectional plugin: the model reads the whole prompt bidirectionally, as it was trained.

Vendor numbers (UnverifiedClaim). A complete local run of the Decision Index 0.3 public suite gives a public index of 67.31 over 37 benchmarks and 140,178 requests, and a raw index of 75.62 (a different metric). The card labels this “self-reported pending maintainer verification” and reports no private or Full score. The X post highlights 96.25 on tool selection and 97.06 on intent classification with out-of-scope detection.

Independent boards (checked 20 Oct). v1.1 is not on Decision Index 0.4 or on JevBench. The board’s only row for the family name is the earlier JEV-27B (DI 0.4 Full 58.55, #47).

Caveats. Hugging Face shows 16,228 downloads in the first days, a high count like the rest of the family, which we keep out of adoption. The X post links a ModelScope mirror, while the card says that mirror is “not currently confirmed”. No score changes; the family page is scored on GEV-26B-Decide.