Skip to content
Metatext Decide — typed decisions for your agents, with a probability on every answerGet early access
Metatext
Sign Up

Winnow

Apache-2.0 Gemma 4 12B fine-tune for local typed decisions, chat and image input, served via /v1/systemone and chat completions.

ServiceresearchApache-2.0
Open source ↗

Overview

Winnow-12B is an EldanRing fine-tune of Gemma 4 12B IT for local typed decisions, ordinary chat and image input. Given a shared state and questions with known options, the Winnow server (built on llama.cpp) scores those options through /v1/systemone without generating an explanation; chat uses /v1/chat/completions on the same loaded model. GGUF downloads include the merged fine-tune in Q8, BF16 and NVFP4 variants, with a projector for vision. The list also mentions a smaller E4B variant.

On the 231-item public JevBench subset, Winnow-12B Q8 scored 85.71 percent, matching hosted Jev 1.13 (both 198 of 231); on a 1,046-item Kev-v9 set it scored 81.55 percent versus 87.00 for Jev. The README notes these panels were used during development. It is a different project from the winnow Claude Code context sieve.

What the decision model decides

TypeQuestionAnswers
ChoiceWhich known option fits this shared state?—

How it works

  1. Download a GGUF (and projector for images).
  2. Run the Winnow server.
  3. Send typed questions to /v1/systemone or chat to /v1/chat/completions.
  4. Read option scores.

Strengths

  • Matches Jev 1.13 on public JevBench in the authors' run.
  • One model for decisions, chat and vision.
  • Apache-2.0, local.

Caveats

  • Trails Jev on the Kev-v9 set.
  • Large downloads (about 12.7 GB for Q8).
  • Evaluation panels were used during development.

At a glance

Type
Service
Status
research
Runtime
llama.cpp-based server (GGUF)
Has code
yes
License
Apache-2.0

Use cases and integrations

How curated lists describe it

  • Apache-2.0 Gemma 4 fine-tunes (12B and E4B) for typed decisions, served from a llama.cpp-based server that answers both `/v1/systemone` and `/v1/chat/completions`.

    AnotiaWang/awesome-decision-models

Frequently asked questions

How does it compare with Jev?

Equal on public JevBench (198/231) but lower on Kev-v9 (81.55 vs 87.00 percent).

Can it chat too?

Yes, via /v1/chat/completions on the same model.

Is it related to the winnow context sieve?

No, it is a separate project with the same name.

Related projects

  • Lichen★ 58

    Drop-in local replacement for Jev that serves /v1/systemone from open weights, including questions about images.

    ServicebetaPython
  • Jev Local

    Local /v1/systemone server in Japanese docs with LFM, Sarashina and fine-tuned ModernBERT-Ja backends, including image input.

    ServicebetaPython
  • NeoHorse-Jev-4B

    Hugging Face card for NeoHorse-Jev-4B: Apache-2.0 4B decision model for text or image-plus-text, served at /v1/systemone.

    ModelbetaPython
  • Standard One

    Standard Thinking's Apache-2.0 3B and 8B decision models on Ministral 3 that score bounded answers via /v1/systemone, with vision.

    ServicebetaPython
  • AutoJev

    Open 27B multimodal decision model built largely by autonomous agents, serving choice, noul and score on a TypeSafe-compatible API.

    ModelresearchPython
  • JEV-27B-VL-Uncensored-GGUF

    Static GGUF quantizations by mradermacher of SargeDev's JEV-27B-VL-Uncensored vision decision model.

    Model