---
title: Gemma4-PRISM-PRO-DQ
canonical_url: "https://www.modelscope.cn/models/Ex0bit/Gemma4-PRISM-PRO-DQ"
md_url: "https://www.modelscope.cn/models/Ex0bit/Gemma4-PRISM-PRO-DQ.md"
repository: Ex0bit/Gemma4-PRISM-PRO-DQ
last_updated: 2026-04-08
license: apache-2.0
pipeline_tag: image-text-to-text
tasks:
  - image-text-to-text
base_model:
  - google/gemma-4-31B-it
base_model_relation: finetune
parameters: 5.5B
tensor_type:
  - BF16
  - U32
library_name:
  - gguf
  - safetensors
  - pytorch
frameworks:
  - pytorch
language:
  - en
downloads: 37
stars: 0
tags:
  - gemma4
  - gemma
  - google
  - gguf
  - zero-refusals
  - prism-dq
  - dynamic-quantization
  - multimodal
  - vision
  - video-text-to-text
  - audio
  - image-text-to-text
  - abliterated
  - text-generation
---

# Gemma4-PRISM-PRO-DQ

> Gemma4-PRISM-PRO-DQ - Ex0bit 在 ModelScope 开源的模型。G4-PRISM-PRO — PRISM Dynamic Quantization

Ex0bit/Gemma4-PRISM-PRO-DQ 是 ModelScope 魔搭社区上的 5.5B 参数image-text-to-text模型，采用 apache-2.0 许可，基于 google/gemma-4-31B-it 构建。

- **Repository**: Ex0bit/Gemma4-PRISM-PRO-DQ
- **License**: apache-2.0
- **Tasks**: image-text-to-text
- **Parameters**: 5.5B
- **Base model**: google/gemma-4-31B-it
- **Tags**: gemma4, gemma, google, gguf, zero-refusals, prism-dq, dynamic-quantization, multimodal, vision, video-text-to-text, audio, image-text-to-text, abliterated, text-generation
- **Downloads**: 37
- **Stars**: 0
- **Last updated**: 2026-04-08

Source: https://www.modelscope.cn/models/Ex0bit/Gemma4-PRISM-PRO-DQ

---

[![Parameters](https://img.shields.io/badge/Parameters-31B-blue)]()
[![Format](https://img.shields.io/badge/Format-GGUF-red)]()
[![Format](https://img.shields.io/badge/Format-MLX-green)]()
[![Quant](https://img.shields.io/badge/Quant-PRISM_Dynamic_(5.02_BPW)-yellow)]()
[![Multimodal](https://img.shields.io/badge/Multimodal-Vision%20%2B%20Text-purple)]()


<div align="center">
  <img src="https://cdn-uploads.huggingface.co/production/uploads/63adf1fa42fd3b8dbaeb0c92/onzfqZmEuOqedGOcyedHc.png" width="800">
</div>

# G4-PRISM-PRO — PRISM Dynamic Quantization

**Gemma 4 31B-It PRISM-PRO-Dynamic-Quant**

- PRISM-PRO: Production model with full over-refusal and bias mechanisms completely removed using State of the Art **PRISM pipeline**.
- DQ: Per-tensor-class mixed-precision allocation derived entirely from weight structure sensitivity analysis - not closed-gated Datasets.

Created by [Ex0bit](https://hf.co/Ex0bit)

---

<div align="center">

### 💡 This model is free for active members, or available for purchase to all others here: https://ko-fi.com/s/0daff0e074. Support My Research & Development efforts. Members Recieve access to the latest PRISM-PRO Model drops on Day-0

[![Ko-fi](https://img.shields.io/badge/Ko--fi-Support%20PRISM-ff5e5b?logo=ko-fi&logoColor=white)](https://ko-fi.com/Ex0bit)

</div>

---

## Model Details

| Property | Value |
|----------|-------|
| Base Model | google/gemma-4-31B-it |
| Architecture | Gemma 4 ISWA (Interleaved Sliding Window Attention) |
| Parameters | 31B dense |
| Quantization | PRISM-PRO-DYNAMIC-QUANT |
| Achieved BPW | 5.02 |
| File Size | ~19 GB |
| Context Length | 262,144 tokens |
| Modalities | Text, Image, Video, Audio |
| Creator | [Ex0bit](https://hf.co/Ex0bit) |

## Supported Modalities

- **Text**: Full instruction-following and chat
- **Image**: Vision understanding via SigLIP encoder (280 tokens/image)
- **Video**: Gemma4VideoProcessor (32 frames, pooled)
- **Audio**: 40ms per token, 750 token sequence length

## PRISM-DQ Quantization

This model uses **PRISM-PRO Dynamic Quantization** — a proprietary per-tensor-class
mixed-precision allocation that assigns different quantization types to different
tensor classes based on 7 structural weight metrics.

Unlike uniform quantization (Q4_K_M, Q5_K_M), PRISM-DQ analyzes each tensor class's
sensitivity to quantization error and allocates precision where it matters most.
Attention projections receive higher precision than FFN layers, with block-level
overrides that protect early layers (high downstream error propagation).

**No calibration data, no importance matrices, no training data required.**

The result: BF16-equivalent quality at 5.02 bits-per-weight — a 68% size reduction
with zero measurable quality loss.

## Usage

### llama.cpp
```bash
llama-server --model Gemma4-PRISM-PRO-DQ-GGUF/gemma4-prism-pro-dq.gguf \
  --port 8080 -ngl 99
```

### LM Studio

Download the GGUF file and load it in LM Studio. The model will be detected as Gemma 4 31B.

### Ollama
```bash
ollama create g4-prism -f Modelfile
```

## Refusal & Bias Removal

This model has been treated to remove bias, over-refusals and propaganda from the base google/gemma-4-31B-it using the State of The Art PRISM pipeline.

## License

Apache 2.0 (inherited from google/gemma-4-31B-it)

## Credits

- Creator: [Ex0bit](https://hf.co/Ex0bit)
- Base model: [Google DeepMind](https://deepmind.google/models/gemma/gemma-4/)
- Quantization engine: PRISM-DQ by [Ex0bit](https://hf.co/Ex0bit)