Winnow — local AI model by EldanRing
Written by Jakub Rusinowski · Last updated
Gemma 4 fine-tunes for typed decisions, shipped as GGUF with a llama.cpp-based server that serves /v1/systemone and chat/vision from one loaded model.
Variants
The smallest Winnow variant needs about 6 GB of VRAM at Q4_K_M — quantized weights plus framework overhead, before any KV cache.
| Model | VRAM at Q4 | VRAM | Context | Run it |
|---|---|---|---|---|
| Winnow 12B → 12B | ~8 GB | Not published | winnow-inference :8091 | |
| Winnow E4B → E4B (8B total) | ~5.6 GB | Not published | winnow-inference :8091 |
Memory is quantized weights plus overhead at Q4_K_M, from the same engine as the GPU & VRAM checker.
How to run Winnow locally
Install Ollama, then pull the tag.
Served by winnow-inference (default port 8091). This model does not run in Ollama.
Pick a size above for its own VRAM figure, speed estimate and install command.
Licence
Commercial use permitted. No usage restrictions beyond attribution.
Applies to: Winnow 12B, Winnow E4BRecommended GPU
The cheapest catalogued GPU that runs Winnow locally (min 6 GB VRAM) is the Intel Arc B570 (10 GB).
Winnow — frequently asked questions
How much VRAM does Winnow need?
Winnow needs about 6 GB VRAM at Q4_K_M quantization for its smallest variant. Variants: Winnow 12B (8 GB, Q4_K_M); Winnow E4B (6 GB, Q4_K_M). On Apple Silicon, unified memory counts toward this requirement.
Can I run Winnow on an RTX 4090 (24 GB)?
Yes — Winnow runs on an RTX 4090 (24 GB) and other 24 GB cards such as the RTX 3090. Smaller variants also fit comfortably on 8–16 GB GPUs at Q4_K_M.
What quantization should I use for Winnow?
Q4_K_M is the best balance of quality and VRAM for Winnow in most cases. Choose Q8_0 for near-lossless quality if you have spare VRAM, or smaller quants (Q3/Q2) only when memory is tight.
How do I run Winnow locally?
Winnow is a decision model: it is called over an HTTP endpoint (/v1/systemone), not chatted with. Where a variant is available in Ollama 0.35 or newer, pull it with `ollama pull` and send requests to the local server; the others ship their own server. Each variant page shows the exact commands for that model.