Winnow-12B

Released
2026-09-20
Last refreshed
2026-10-15
Status
Researched 36d ago
Open sourceCommercial use: permittedMultimodalAgentsVisionJSON / Tool useAgent

Winnow-12B is available now for agents, vision, and json / tool use with open-source; evaluate it while provider pricing coverage matures.

EldanRing releases · 1 in the last 12 monthsChangelog →
Specifications
Family
Winnow
Released
2026-09-20
Context
64K
Parameters
12B
Openness
Open source
License
Apache 2.0OSI-approvedPermitted
Weights
Available
Code
Unknown
Training
Fine-tuned
Pricing

No tracked provider pricing is available yet.

About

Winnow-12B is EldanRing's local decision model, a rank-32 LoRA fine-tune of Gemma 4 12B IT merged into the weights and released as GGUF files (BF16, Q8_0, and NVFP4) with an optional F16 vision projector. It answers typed choice, yes/no (called "noul" in the API) or score questions against a shared state by prefilling the state once, forking question branches, and reading answer-token logits without generating text. The same loaded model also serves ordinary chat, streaming, and image input. Its llama.cpp-based inference server exposes /v1/systemone and /v1/chat/completions.

Providers

No active provider routes are tracked for this model yet.

Capabilities

VisionMultimodalStructured Outputs

Benchmark peer barsfor Agents

No task-mapped benchmark peers are available for this model yet.

Benchmark scores(2)

Scores are benchmark-specific and are direction-aware: the same numeric gap can mean very different outcomes across suites. Use the leaderboard context and this model's provider route to decide whether the winning margin is meaningful for your workload.
BenchmarkScoreVersionEvaluationSource
68.9v1.6.0Observed 2026-10-02—
49.30.3.2Observed 2026-10-17—

Migration checks

No linked migration route is available for this model yet.

API versions

EldanRing/Winnow-12BWinnow-12BWinnow-12B Q8winnow-12b