---

base_model: coder3101/Qwen3.5-9B-heretic

base_model_relation: quantized

library_name: gguf

license: apache-2.0

language:

• multilingual

pipeline_tag: image-text-to-text

tags:

• gguf
• text-generation
• multimodal
• vision
• reasoning
• thinking
• agentic
• tool-use
• abliterated
• uncensored
• qwen
• q4_k_m
• conversational

---

Qwen3.5-9B-heretic — GGUF

This repository hosts GGUF weights and the associated vision projection matrix for Qwen3.5-9B-heretic, quantized from the source floating-point tensors provided by coder3101/Qwen3.5-9B-heretic.

About the Model

Qwen3.5-9B is a vision-language model by the Qwen Team (Alibaba) with a hybrid architecture combining Gated Delta Networks and sparse Mixture-of-Experts (MoE) for high-throughput inference with minimal latency. The model supports 201 languages and dialects, integrates a native vision encoder via early-fusion multimodal training, and was post-trained with reinforcement learning at scale across million-agent environments. It natively supports tool calling, agentic pipelines, and extended context up to 262,144 tokens. The model thinks by default, generating reasoning content before its final response.

The heretic suffix denotes post-processing via Heretic v1.2.0 with the Multi-directional refusal suppression method performed by coder3101, which simultaneously targets multiple refusal directions across model layers to achieve a thorough and stable abliteration.

---

Model & Architecture Specifications

| Property | Value |

|---|---|

| Base Architecture | Qwen3.5 hybrid (Gated Delta Networks + sparse MoE) |

| Developed by | Qwen Team (Alibaba) |

| Parameters | 9B language backbone (~10B total with vision encoder) |

| Primary Use | Multimodal reasoning, coding, tool use, agentic tasks |

| Context Window | 262,144 tokens (extensible to ~1M) |

| Vision Encoder | Integrated (early-fusion) |

| Languages | 201 languages and dialects |

| Abliteration Tool | Heretic v1.2.0 |

| Abliteration Method | Multi-directional refusal suppression |

| Prompt Format | ChatML |

---

Abliteration Parameters

| Parameter | Value |

|---|---|

| direction_index | 14.33 |

| attn.o_proj.max_weights | [0: 1.04, 1: 1.49, 2: 1.07, 3: 1.31] |

| attn.o_proj.max_weight_position | 20.19 |

| attn.o_proj.min_weights | [0: 0.06, 1: 0.73, 2: 0.17, 3: 0.54] |

| attn.o_proj.min_weight_distance | 16.54 |

| mlp.down_proj.max_weights | [0: 1.47, 1: 0.82, 2: 1.06, 3: 1.24] |

| mlp.down_proj.max_weight_position | 20.06 |

| mlp.down_proj.min_weights | [0: 0.55, 1: 0.13, 2: 0.57, 3: 1.00] |

| mlp.down_proj.min_weight_distance | 11.39 |

Abliteration Performance

> [!NOTE]

> The metrics below are self-reported by the original model author (coder3101) and have not been independently reproduced.

| Metric | This model | Original (Qwen/Qwen3.5-9B) |

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

| KL divergence | 0.1334 | 0 (by definition) |

| Refusals | 10/100 | 86/100 |

---

Quantization Details

| Property | Value |

|---|---|

| Text Tensors | Q4_K_M |

| Vision Tensors | BF16 (unquantized, preserves visual feature quality) |

Quantization Command

./llama-quantize /content/model-bf16.gguf \
  /content/Qwen3.5-9B-heretic-Q4_K_M.gguf q4_k_m

---

Repository Files

| Filename | Format | Size | llama.cpp Build | Description |

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

| Qwen3.5-9B-heretic-Q4_K_M.gguf | Q4_K_M | 5.63 GB | b9803 | Main model weights |

| mmproj-Qwen3.5-9B-heretic-BF16.gguf | BF16 | 922 MB | b9803 | Vision projector — required for all image inputs |

---

Inference

> [!IMPORTANT]

> The vision projector (--mmproj) must be supplied at runtime whenever image inputs are used. Omitting it disables multimodal capability entirely.

> [!NOTE]

> This model thinks by default and will emit reasoning content before its final answer. To disable this, set enable_thinking: False if your inference stack supports it; standard llama.cpp does not expose this toggle directly, so expect reasoning output unless you strip it client-side.

llama.cpp CLI (with image)

./llama-cli \
  -m Qwen3.5-9B-heretic-Q4_K_M.gguf \
  --mmproj mmproj-Qwen3.5-9B-heretic-BF16.gguf \
  -c 8192 \
  -ngl 99 \
  --image "path/to/image.jpg" \
  -p "<|im_start|>user\nDescribe what you see and answer any questions about the image.<|im_end|>\n<|im_start|>assistant\n"

OpenAI-Compatible API Server

./llama-server \
  --host 0.0.0.0 \
  --port 8080 \
  -m Qwen3.5-9B-heretic-Q4_K_M.gguf \
  --mmproj mmproj-Qwen3.5-9B-heretic-BF16.gguf \
  -c 16384 \
  -ngl 99 \
  --flash-attn

---

Prompt Format (ChatML)

<|im_start|>system
You are a helpful multimodal assistant.<|im_end|>
<|im_start|>user
Your question or image payload here.<|im_end|>
<|im_start|>assistant

---

Notes & Limitations

• This model is abliterated and will generate content that standard aligned models refuse. Use responsibly and in compliance with applicable laws.
• Vision tensors are kept at BF16 to prevent spatial feature degradation during visual token extraction.
• For better output quality at this quantization level, consider the imatrix variant in the companion repository.