Winnow Models by EldanRing

EldanRingApache 2.0Open sourceAgent
EldanRing releases · 1 in the last 12 monthsChangelog →
1 model2026Up to 64K ctx

Last refreshed 2026-10-15. Next refresh: weekly.

Details

ResearcherEldanRing
LicenseApache 2.0OSI-approved
Commercial usePermitted
Models1
Released2026
Max context64K

Capabilities

VisionAll models
MultimodalAll models
Structured OutputsAll models

About

Winnow is EldanRing's family of open-weight local decision models built on Gemma 4. Each answers typed choice, yes/no (called "noul" in the API) or score questions against a shared state by reading answer-token logits without generating text, and also serves ordinary chat and image input from the same loaded model through a llama.cpp-based server with /v1/systemone and /v1/chat/completions. Released September 2026 as GGUF files under Apache 2.0.

Decision facts

Best fit
agentvision and multimodal workstructured outputs
Capability starting point
Winnow-12B with 64K context and structured outputs and multimodal inputs
Lowest tracked input
Not tracked
Closest related family
Quyet

Current Variants

Use-when guidance is based on each model's tracked capabilities, context window, release date, and replacement status.

1 in view
Winnow-12BCurrent

Use when the workload needs 64K context, 12B parameters, and structured outputs.

2026-0964K context12B parametersstructured outputs

Release Timeline

1 release group
2026-09
1 current
Winnow-12B
64K context12B parametersstructured outputs
Current

Specifications(1 models)

Winnow model specifications comparison
ModelReleasedContextParametersVisionMultimodalStructured Outputs
Winnow-12B2026-0964K12BYesYesYes