Winnow-12B
A Gemma 4 12B fine-tune for typed decisions that also chats and reads images. Ships as GGUF for llama.cpp, holds a 64K context on a 16 GB GPU, and serves /v1/systemone next to /v1/chat/completions.
On JevBench's public subset (231 items) the Q8_0 build scores 85.71%, the same as Jev 1.13 measured through OpenRouter on the same inputs, and 81.55% on Kev v9 clean (Jev 87.00%). It is a rank-32 LoRA merged into Gemma 4 12B IT and shipped as BF16 and Q8_0 GGUF with an optional vision projector; on an RTX 5070 Ti the Q8 build holds a 64K context with an image in 15 GiB, and a cached four-question request takes 143 ms. Probabilities use temperature 1.0 with no fitted calibration map. Sibling: Winnow-E4B on Gemma 4 E4B, 80.52% on the same JevBench subset at about twice the decisions per second. Both are Apache-2.0, like Gemma 4.
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 calibrated probability
- classify — assigns a category from a fixed taxonomy
- route — sends the input to one of several destinations
At a glance
| Parameters | 12B |
| Base model | google/gemma-4-12b-it |
| Maker | EldanRing |
| Released | 2026-09-20 |
| License | apache-2.0 |
| Reported accuracy | 85.7% |
| Reported latency | 143 ms for a cached four-question request on an RTX 5070 Ti (Q8) |
Get the weights
pip install systemonemodels
systemone pull eldanring/winnow
The files are served from the maker's Hugging Face repository, EldanRing/Winnow-12B, and verified against the checksums recorded here.