AUBIN-12B
Personal open-weights project: LoRA adapters on Gemma 4 (12B here; 31B and E4B siblings) that answer typed choice, yes/no and score questions by reading option-letter logits with a fitted temperature. Uncertain questions get a short written rationale and a re-score.
Jev-inspired. A rank-16 LoRA on google/gemma-4-12B-it with a fitted temperature; it needs the Gemma 4 base, and the repo holds the adapter and the maker's aubin package, which also serves a Laya/Jev-style endpoint. It scores option letters in one forward pass, but by default writes a short rationale and re-scores when the top two options are close (26 to 32% of questions, per the maker), so it is not purely one-pass unless that step is turned off. Siblings AUBIN-31B and AUBIN-E4B-v3. The family's web-agent, screenshot-click and grid-control adapters and the AUBIN-Omni router are not decision models and are not listed. On LocalLLaMA/typed-decisions the maker reports 72.9% for this adapter alone; the card's 77.55% headline fuses it with a separate classifier.
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 | 12B |
| Base model | google/gemma-4-12B-it |
| Maker | Emre Vergili (Norovox) |
| Released | 2026-09-29 |
| License | apache-2.0 |
| Reported accuracy | 87.8% |
| Reported latency | 486 ms per question on the fast path, 4-bit on one T4; about 6 to 7 s more on questions it thinks about |
Get the weights
pip install systemonemodels
systemone pull emrevrg/aubin
The files are served from the maker's Hugging Face repository, emrevrg/AUBIN-12B, and verified against the checksums recorded here.