GEV-26B-Decide
AutoTrust's decision model on Gemma-4-26B-A4B (about 4B parameters active per token). A LoRA and a 24-slot head give a probability for every option in one pass over text or images; opt-in adaptive thinking lets the unmodified base reason when System 1 is unsure.
Noul, Score (0 to 5) and Choice over 2 to 256 options (more than 16 are read in groups of 16, then a final round), served by a patched vLLM development build with POST /v1/decide; it needs one GPU with 80 GB or more. The repository holds google/gemma-4-26B-A4B-it unchanged (System 2) plus the System 1 LoRA and head. With thinking set to "auto", a question whose leading option is below 0.8 goes to the base model's thinking mode, which costs seconds per question. Image decisions are zero-shot, because the head was trained on text. Same weights as autotrust/JEV-Gemma4-26B-A4B, its earlier name; an NVFP4 build exists. The card does not name the System 1 teacher; its JEV siblings are distilled from TypeSafe Jev 1.13 outputs, and AutoTrust states no affiliation with TypeSafe, so check TypeSafe's terms before mirroring. Training used the public training splits of some Decision Index datasets (BANKING77, CLINC150). AutoTrust reports 83.4% with adaptive thinking on its six-set suite. Its own scoring of the third-party Decision Index 0.2.1 suite gives 58.05 for System 1 alone and 62.48 with adaptive thinking on Knowledge & Reasoning; neither is a board result. Weights Apache-2.0; the base is under Google's Gemma 4 terms.
What it decides
- choice — picks one option from a set
- score — places the input on an ordered scale
- noul — answers a yes/no question with one probability
At a glance
| Parameters | 26B |
| Base model | google/gemma-4-26b-a4b-it |
| Maker | AutoTrust AI Lab |
| Released | 2026-10-02 |
| License | apache-2.0 |
| Reported accuracy | 73.3% |
| Reported latency | 45 ms median per System 1 decision on one B200 (vLLM); adaptive thinking adds about 1 s per 250 thinking tokens |
Get the weights
pip install systemonemodels
systemone pull autotrust-ai/gev-26b-decide
The files are served from the maker's Hugging Face repository, , and verified against the checksums recorded here.