Winnow 12B

• License: Apache-2.0
• Original model: EldanRing/Winnow-12B
• Base model: google/gemma-4-12B-it
• Served modalities: Text and image input; choice, score, and noul output.

Description

Winnow 12B is EldanRing's merged Gemma 4 fine-tune for typed decisions, chat, and image input. This registry entry serves typed decisions through ggmlc-custom-decider with protocol=winnow. The engine uses Winnow's Gemma prompt and next-token option-label probabilities, not the Decider prompt. Each question is evaluated independently; shared-prefix caching is not implemented.

Run

self serve winnow:12b

Use the Decision API at /v1/systemone. Image input uses the request's images array; the current adapter accepts one image per request. Chat, audio, and video are not exposed by this entry.

Quantizations and Files

The default and only registered variant is q8, using the author's merged gguf/Winnow-12B-Q8_0.gguf (12,669,646,592 bytes) and matching F16 vision projector gguf/mmproj-Winnow-12B.gguf (175,115,840 bytes). Both files are GGUF, pinned to revision 859a2cb24921d8b579badce6caa469f9290f3bc1 with SHA-256 checks in the registry. No separate base model, LoRA adapter, or speculative assistant is needed.

Settings and Limits

Startup context is 8,192 tokens; the upstream Q8 runtime profile has tested 65,536 positions, including image tokens. Override with --set context_size=65536 when memory permits. Decisions do not generate text. Model download size excludes KV cache and runtime overhead.

The upstream default decision temperature of 1.0 applies to option-logit scoring. Up to 64 options are supported. On 2026-10-13, self check passed all CPU choice-routing, option-order, true/false, and score-ordering probes. Image-decision tests also passed red/green/blue classification, malformed-image rejection, and rejection of requests with more than one image. These smoke tests are not a benchmark or an upstream numerical-parity check.

Quality benchmarks

Decision Index: #47 (49.27)

EldanRing/Winnow-12B

Published result for the named upstream model and evaluated configuration; checkpoint revision and local quantization equivalence are not guaranteed.

Evaluated: winnow-12b-q8 (Q8_0)

Edition: Decision Index 0.3.2

Checked 2026-10-13

JevBench: #11 (68.89)

EldanRing/Winnow-12B

Published result for the named upstream model and evaluated configuration; checkpoint revision and local quantization equivalence are not guaranteed.

Evaluated: Winnow-12B Q8

Edition: v1.6.1

Checked 2026-10-13

Typed Decisions: Publisher-reported

EldanRing/Winnow-12B

Winnow publisher comparison, Q8 weights and Q8 KV, 400 cases / 2000 decisions. Teacher agreement, not correctness. Development exposure disclosed; not an official leaderboard rank.

Dataset	Metric	Score
LocalLLaMA/typed-decisions (test)	Teacher agreement (%)	70.00

Benchmarks

No submitted benchmark reports yet.