Winnow-12B

Base model: Gemma 4 12B IT by Google DeepMind
Source: https://huggingface.co/google/gemma-4-12B-it
Source revision: 707f0a3b8a3c7ad586ed01e27eafbad8a27dd0f7
Upstream license: Apache License, Version 2.0
License reference: https://ai.google.dev/gemma/docs/gemma_4_license

Winnow modifications are developed by EldanRing. The released model weights
modify the base through task-specific LoRA fine-tuning and adapter merging.
The Q8_0 GGUF is converted and quantized from those merged weights. Configuration
and weight sharding were updated during export; tokenizer and chat template
assets were preserved. The matching vision projector is unchanged by the
decision fine-tuning. Artifact hashes are recorded in SHA256SUMS and
release-manifest.json.

This file records Winnow's attribution and modifications. It does not imply
that the upstream model distribution supplied a NOTICE file.

The separate Winnow inference code builds on llama.cpp by Georgi Gerganov and
contributors under the MIT license. Its repository contains the applicable
copyright and license text. No llama.cpp source or binary is included here.

Winnow is an independent project; no upstream endorsement is implied.