README.md

---
license: mit
widget:
- text: a conversation between cgg and connector
  output:
    url: output-sample.mp4
tags:
- cow
- gguf-node
---

# 🐮cow architecture gguf encoder
- don't need to rebuild tokenizer from metadata ⏳🥊
- don't need separate tokenizer file 🐱🥊
- no more oom issues (possibly) 💫💻🥊

## eligible model example
- use **cow-mistral3small** [7.73GB](https://huggingface.co/chatpig/flux2-dev-gguf/blob/main/cow-mistral3-small-q2_k.gguf) for [flux2-dev](https://huggingface.co/chatpig/flux2-dev-gguf)
- use **cow-gemma2** [2.33GB](https://huggingface.co/calcuis/cow-encoder/blob/main/cow-gemma2-2b-q4_0.gguf) for [lumina](https://huggingface.co/calcuis/lumina-gguf)
- use **cow-umt5base** [451MB](https://huggingface.co/calcuis/cow-encoder/blob/main/cow-umt5base-iq4_nl.gguf) for [ace-audio](https://huggingface.co/calcuis/ace-gguf)
- use **cow-umt5xxl** [3.67GB](https://huggingface.co/calcuis/cow-encoder/blob/main/cow-umt5xxl-q2_k.gguf) for [wan-s2v](https://huggingface.co/calcuis/wan-s2v-gguf) or any wan video model

![screenshot](https://raw.githubusercontent.com/calcuis/comfy/master/cow-encoder.png)

the example workflow above is from wan-s2v-gguf; cow encoder is a special designed clip, even the lowest q2 quant still working very good; upgrade your node for cow-encoder support🥛🐮 and do drink more milk

![screenshot](https://raw.githubusercontent.com/calcuis/comfy/master/stars.jpg)

<Gallery />

### **reference**
- gguf-node ([pypi](https://pypi.org/project/gguf-node)|[repo](https://github.com/calcuis/gguf)|[pack](https://github.com/calcuis/gguf/releases))