GEMMA 4 E4B ABLITERATED
A ComfyUI-compatible repack of wangzhang's abliterated Gemma 4 E4B weights, formatted as a single safetensors file usable with ComfyUI's standard Load CLIP and TextGenerate nodes.
WHAT THIS IS
This file combines:
- The abliterated language model weights from wangzhang/gemma-4-E4B-it-abliterix
https://huggingface.co/wangzhang/gemma-4-E4B-it-abliterix
- The vision encoder, audio encoder, multimodal projector, and tokenizer components from Comfy-Org/gemma-4 (unmodified)
https://huggingface.co/Comfy-Org/gemma-4
The result is a drop-in replacement for ComfyUI's standard Gemma 4 E4B text encoder file, with the language model behavior of the upstream abliterated weights.
WHAT THIS IS NOT
This is not a new model. The intelligence, capabilities, and modifications were done upstream by Google (base model) and wangzhang (abliteration). This file is a format repack it puts existing modified weights into ComfyUI's expected single-file structure so users can load them with standard nodes.
CREDITS
Base model: Google DeepMind — Gemma 4 E4B IT
https://huggingface.co/google/gemma-4-E4B-it
Abliteration work: wangzhang — Abliterix method
https://huggingface.co/wangzhang/gemma-4-E4B-it-abliterix
ComfyUI integration and reference format: Comfy-Org / Kijai
https://huggingface.co/Comfy-Org/gemma-4
INSTALLATION
Place the .safetensors file in:
ComfyUI/models/text_encoders/
Use with ComfyUI's standard Load CLIP and TextGenerate nodes by selecting this file as the text encoder.
TECHNICAL NOTES
The repack reconstructs Gemma 4's shared-KV layer structure: layers 24-41 share key and value projections with layer 23, requiring tensor cloning during conversion. The k_norm parameters for these shared layers are sourced from the reference Comfy-Org file to maintain compatibility with ComfyUI's loader.
Vision encoder (4 layers, 768 hidden), audio encoder (12 layers), and multimodal projector are bundled unchanged from the reference file. Tokenizer is the standard Gemma 4 tokenizer.
LICENSE
This model is provided under the Gemma Terms of Use:
https://ai.google.dev/gemma/terms
Use of this model is also subject to Google's Gemma Prohibited Use Policy:
https://ai.google.dev/gemma/prohibited_use_policy
By downloading this model, you agree to be bound by these terms. You are responsible for your use of the model and any content it generates.
RESPONSIBLE USE
This is an abliterated model with reduced refusal behavior on the base language model. The vision and audio encoders are unmodified.
The user is responsible for ensuring their use complies with applicable laws and the Gemma Prohibited Use Policy. The author of this repack does not endorse generating content that would violate those terms.
Description
FAQ
Comments (5)
THX! Any chance to update base model knowledge edge? Now, june 2026, but it stuck on jan 2025 and doesnt know about many actual AI models at all.
User error/ignorance.
Used it to analyze a couple of pictures. Doesn't have much of a sense of female nudity. Thinks every picture is an intense breathtaking work of art. Has no concept of a penis (thinks it's a finger, or a 'touch' or just doesn't exist).
Try giving it deliberate instructions to recognize those things via the text generate node. Wire your instructions into the box on that node. I used AI to build the instructions for the specific model type I'm using.
I've also found that giving it some basic context helps. I use a simple string like context1 and perform a text replace. Outside of the instructions for prompt formatting, I create a section like: The user will provide you with basic context. Use this context as a baseline and starting point for generating the prompt. Here is the user's context: context1
Then the text replace finds the basic string and places your context into the instructions. That way, the model has instructions for formatting and your context to generate the prompt. It's a process but it works.
@Matthew3179 I stand corrected. Thank you. I already had a very complex set of comments for getting to describe what I need the output to look like. Now I've added my own description to the start... And it seems to work. Still wants to make things a much warmer sensual image however. I'll need to refine. THanks for the guidance.
Any chance you can do the 12b version that came out recently?