Winnow 12B — VRAM & /v1/systemone setup
Written by Jakub Rusinowski · Last updated
EldanRing's 12B decision model on Gemma 4 12B, shipped as GGUF and served by the llama.cpp-based winnow-inference server, which also handles chat and vision.
Winnow 12B needs about 8 GB of VRAM at Q4_K_M — quantized weights plus framework overhead, before any KV cache. On Apple Silicon that figure comes out of unified memory.
Call Winnow 12B
Served by winnow-inference (default port 8091). This model does not run in Ollama.
- Question types
- choice noul (yes / no) score
- Inputs
- text image
- Serve with
winnow-inference :8091- Endpoint
/v1/systemone
http://localhost:8091/v1/systemone
Setup for servers other than Ollama →
Build a request for this model Decision models guide
Other decision models: Nimble 9B · Tev1 4B · Tev1 0.8B · Winnow E4B · Decider 2B · Decider 4B · Decider 35B-A3B (NVFP4) · Decider 12B Vision · Decider 31B v2 (NVFP4) · JevK5 4B · Intern-Decision 4B · AutoJev 27B · Laya · Clef 27B · Clef-Flash 9B · Rune 26B-A4B v3 · OpenDecider-nano · OpenDecider-small · OpenDecider-small MLX 4-bit · Bosun v3.1 0.6B · Bosun v3.1 1.7B · d1-3B · d1-omni-600M · THX-01 322M · mammoth-1 27B · elephant-1 9B · Strands Decider 2B (hobson-v21) · kev 4B · kev 9B · OpenJev 27B
Hardware fit
Weights plus overhead plus the KV cache at a 8,192-token prompt, on NVIDIA RTX 4090 (24 GB). A publisher build is sized from its file; the other rows are modelled at a standard quant. Decision requests are short, so no long-context figure is shown. Image tokens are not modelled: each image adds to the prompt, so a request with images needs more memory than these rows.
| Quant | Memory | VRAM | Fit |
|---|---|---|---|
| BF16 Publisher build · 23.83 GB file | 25.9 GB | Offload | |
| Q4_K_M 4.83 bpw · modelled quant | 9.3 GB | Fits | |
| Q6_K 6.56 bpw · modelled quant | 11.9 GB | Fits | |
| Q8_0 8.50 bpw · modelled quant | 14.8 GB | Fits |
Published file size: Q8_0 12.67 GB · BF16 23.83 GB. A download size from the model publisher — not a VRAM requirement.
Runs in the Winnow llama.cpp-based server (default port 8091): /v1/systemone + /v1/chat/completions + vision. Tested: RTX 5070 Ti 16 GB (Q8_0, 64K ctx); Apple Silicon 24 GB+. Not usable on Ollama's /v1/systemone.
How it was scored
Two different suites on two different scales. Never compare the numbers across the two cards.
Decision Index 0.2.1 (snapshot 2026-09-28)
Decision Index 0.2.1Community Decision Index (not by the model's authors)- ECE (lower is better)
- 0.1679
- Median compute
- 72.5 ms on 1x NVIDIA RTX PRO 6000 (96 GB)
Community-maintained; not affiliated with the model authors.
Source ↗Decision models are not ranked on chat, creative or coding scores.
Specifications
Verified — Checked against the primary source — the model card or the vendor spec page — and corroborated by a second independent source.
- Parameters
- 11.96B
- Context window
- Not published
- Architecture
- Fine-tune of google/gemma-4-12B
- Provider
- EldanRing
- Licence
- Apache-2.0
- Practical deployment
- Laptop
- Specified at
- Q4_K_M
- System RAM
- 18 GB
- Record updated
- 2026-09-30
Commercial use permitted. No usage restrictions beyond attribution.
Other Winnow sizes
Winnow 12B — frequently asked questions
What is Winnow 12B?
Winnow 12B is a decision model: you send it a state and typed questions (choice, yes/no/unknown, or a score) and it returns one answer per question with a probability for every option. It is not a chat model.
How do I run Winnow 12B locally?
Winnow 12B does not run in Ollama. Served by winnow-inference (default port 8091). This model does not run in Ollama. See the setup guide for servers other than Ollama.
How much memory does Winnow 12B need?
About 8 GB for the weights plus overhead at Q4_K_M, before the prompt's KV cache. Decision prompts are short, so the cache stays small.
How accurate is Winnow 12B?
Its published score is on the community Decision Index 0.2.1, which its authors did not build. It is not a calibration guarantee: test on your own examples.