---

language:

- en

license: gemma

tags:

- gemma4

- gguf

- quantized

- 31b

base_model: google/gemma-4-31b-it

pipeline_tag: text-generation

---

Gemma-4-31B-JANG_4M-CRACK-GGUF

GGUF quantizations of Gemma-4-31B-JANG_4M-CRACK for use with llama.cpp, LM Studio, Ollama, and other GGUF-compatible inference engines.

About the Model

• Base model: google/gemma-4-31b-it
• Architecture: Gemma 4 Dense Transformer (31B parameters, 60 layers)
• Features: Hybrid Sliding/Global Attention, Vision + Audio multimodal
• Modification: CRACK abliteration (refusal removal) + JANG v2 mixed-precision quantization

Why This Conversion?

The original model uses JANG v2 mixed-precision MLX quantization (attention 8-bit + MLP 4-bit), which is only compatible with vMLX. Standard tools (llama.cpp, LM Studio, oMLX, mlx-lm) cannot load this format due to mixed per-layer bit widths.

This repository provides standard GGUF quantizations that work everywhere.

Conversion Process

Original (JANG v2 MLX safetensors, ~18GB)
    ↓ dequantize (attention 8-bit → f16, MLP 4-bit → f16)
Intermediate (float16 safetensors, ~60GB)
    ↓ convert_hf_to_gguf.py + quantize
GGUF (various quantizations)

Note: Since the original was already quantized (avg 5.1 bits), the dequantized f16 intermediate is an approximation. Re-quantizing to GGUF introduces minimal additional quality loss since the attention layers were preserved at 8-bit in the original.

Available Quantizations

| File | Quant | Size | Quality | Notes |

|------|-------|------|---------|-------|

| gemma-4-31b-jang-crack-Q3_K_M.gguf | Q3_K_M | ~14 GB | Acceptable | Minimum viable quality |

| gemma-4-31b-jang-crack-Q4_K_M.gguf | Q4_K_M | ~18 GB | Good | Best size/quality balance |

| gemma-4-31b-jang-crack-Q5_K_M.gguf | Q5_K_M | ~21 GB | Better | Recommended if RAM allows |

| gemma-4-31b-jang-crack-Q6_K.gguf | Q6_K | ~25 GB | Very Good | High quality |

| gemma-4-31b-jang-crack-Q8_0.gguf | Q8_0 | ~33 GB | Near lossless | Closest to original |

System Requirements

| Quantization | Minimum RAM | Recommended |

|-------------|------------|-------------|

| Q3_K_M | 20 GB | 24 GB |

| Q4_K_M | 24 GB | 32 GB |

| Q5_K_M | 28 GB | 36 GB |

| Q6_K | 32 GB | 40 GB |

| Q8_0 | 40 GB | 48 GB |

Usage

LM Studio

Download any .gguf file and open it in LM Studio.

llama.cpp

./llama-cli -m gemma-4-31b-jang-crack-Q4_K_M.gguf -p "Hello" -n 256

Ollama

echo 'FROM ./gemma-4-31b-jang-crack-Q4_K_M.gguf' > Modelfile
ollama create gemma4-crack -f Modelfile
ollama run gemma4-crack

License

Gemma License

Disclaimer

This model has had safety guardrails removed. Use responsibly and in compliance with applicable laws.