AJev (Gemma 4 26B-A4B, lora1)
An open Jev-style decision model by Andy Zhang: a LoRA adapter on Gemma 4 26B-A4B (also on Gemma 4 12B) that reads a state and returns a temperature-calibrated probability for every option of a yes/no, choice or score question in one forward pass.
The repo holds a rank-32 LoRA (about 149 MB) and per-type temperatures, not the base weights (about 55 GB in bf16). Sibling ajev-gemma4-12b-lora7; the older lora5 is kept for leaderboard reproduction. The maker's own held-out runs: 0.887 on JevBench's public items (a third-party benchmark), 0.790 on Kev's transfer set, 0.780 on typed-decisions. His own run of the third-party Decision Index 0.2.1 gives 57.42 (not reproduced by the maintainers; the board lists lora5 at 50.02). Trained in part on training splits of benchmarks related to that board, with decontamination stated. The cards say some training data carries non-commercial or attribution terms. Inference code (cmzy/ajev-infer, Apache-2.0) serves /v1/systemone. Not affiliated with TypeSafe AI.
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 | Andy Zhang |
| Released | 2026-10-05 |
| License | apache-2.0 |
| Reported latency | 49 ms median for sequential requests on one RTX PRO 6000 |
Get the weights
pip install systemonemodels
systemone pull andy-zhang/ajev
The files are served from the maker's Hugging Face repository, andyzhang232/ajev-gemma4-26b-a4b-lora1, and verified against the checksums recorded here.