---
library_name: transformers.js
base_model:
- huihui-ai/Huihui-gemma-4-E2B-it-abliterated
pipeline_tag: image-text-to-text
tags:
- abliterated
- uncensored
- gemma4
- onnx
- webgpu
license: apache-2.0
---

# Huihui-gemma-4-E2B-it-abliterated (ONNX)

ONNX export of [huihui-ai/Huihui-gemma-4-E2B-it-abliterated](https://huggingface.co/huihui-ai/Huihui-gemma-4-E2B-it-abliterated)
(uncensored / abliterated version of google/gemma-4-E2B-it) for use with
[Transformers.js](https://huggingface.co/docs/transformers.js) in the browser via WebGPU.

**Live demo:** [kermelp/Huihui-gemma4-abliterated-webgpu](https://huggingface.co/spaces/kermelp/Huihui-gemma4-abliterated-webgpu)

## Usage with Transformers.js

```js
import { AutoProcessor, Gemma4ForConditionalGeneration, TextStreamer } from "@huggingface/transformers";

const model_id = "kermelp/Huihui-gemma-4-E2B-it-abliterated-ONNX";
const processor = await AutoProcessor.from_pretrained(model_id);
const model = await Gemma4ForConditionalGeneration.from_pretrained(model_id, {
  dtype: "q4f16",
  device: "webgpu",
});

const messages = [
  {
    role: "user",
    content: [{ type: "image" }, { type: "text", text: "Describe this image." }],
  },
];
const prompt = processor.apply_chat_template(messages, {
  enable_thinking: false,
  add_generation_prompt: true,
});
```

## Conversion

Converted from bf16 safetensors with `optimum-cli export onnx --task image-text-to-text`,
then quantized to 4-bit weights + fp16 activations (`q4f16`) with `onnxruntime` `MatMulNBitsQuantizer`.

> ⚠️ This model has had its refusal behavior removed via abliteration. It will answer
> prompts the original Gemma 4 refuses. For research and evaluation only.