---

base_model: coder3101/QwenPaw-Flash-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
• agentic
• tool-use
• memory
• abliterated
• uncensored
• qwen
• q4_k_m
• iq4_nl
• imatrix
• conversational

---

QwenPaw-Flash-9B-heretic — Importance Matrix GGUF

This repository hosts importance-matrix (imatrix) optimized GGUF weights, available in multiple quantization formats, and the associated vision projection matrix for QwenPaw-Flash-9B-heretic, quantized from the source floating-point tensors provided by coder3101/QwenPaw-Flash-9B-heretic.

About the Model

QwenPaw-Flash-9B is a model by agentscope-ai, fine-tuned from Qwen3.5-9B and deeply optimised for the QwenPaw autonomous agent scenario. Since its training phase, the model has been specifically refined for QwenPaw tasks, delivering enhanced agentic performance in tool invocation, command execution, memory management, and multi-step planning. It inherits the base model's hybrid Gated Delta Network + sparse MoE architecture, along with vision capability carried over from the Qwen3.5-9B base.

The heretic suffix denotes post-processing via the Heretic v1.2.0 Arbitrary-Rank Ablation (ARA) method with row-norm preservation performed by coder3101, ensuring refusal responses never interrupt agent workflows or autonomous task loops.

---

Model & Architecture Specifications

| Property | Value |

|---|---|

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

| Developed by | agentscope-ai |

| Fine-tuned from | Qwen/Qwen3.5-9B |

| Primary Use | Autonomous agents, tool calling, memory management, multi-step planning |

| Context Window | 262,144 tokens |

| Vision Capability | Inherited from Qwen3.5-9B (early-fusion) |

| Abliteration Tool | Heretic v1.2.0 |

| Abliteration Method | Arbitrary-Rank Ablation (ARA) with row-norm preservation |

| Prompt Format | ChatML |

---

Abliteration Parameters

| Parameter | Value |

|---|---|

| start_layer_index | 13 |

| end_layer_index | 16 |

| preserve_good_behavior_weight | 0.9042 |

| steer_bad_behavior_weight | 0.0003 |

| overcorrect_relative_weight | 1.0109 |

| neighbor_count | 13 |

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 (agentscope-ai/QwenPaw-Flash-9B) |

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

| KL divergence | 0.0099 | 0 (by definition) |

| Refusals | 12/100 | 96/100 |

---

Quantization Details

| Property | Value |

|---|---|

| Text Tensor Types | Q4_K_M, IQ4_NL (both with imatrix calibration) |

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

| Importance Matrix | bartowski's calibration_datav5.txt |

Quantization Commands

# Q4_K_M
./llama-quantize --imatrix /content/imatrix.dat \
  /content/model-bf16.gguf \
  /content/QwenPaw-Flash-9B-heretic-Q4_K_M-imatrix.gguf q4_k_m

# IQ4_NL
./llama-quantize --imatrix /content/imatrix.dat \
  /content/model-bf16.gguf \
  /content/QwenPaw-Flash-9B-heretic-IQ4_NL-imatrix.gguf iq4_nl

---

Repository Files

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

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

| QwenPaw-Flash-9B-heretic-Q4_K_M-imatrix.gguf | Q4_K_M | 5.63 GB | b9821 | Main model weights, imatrix-calibrated |

| QwenPaw-Flash-9B-heretic-IQ4_NL-imatrix.gguf | IQ4_NL | 5.42 GB | b9843 | Main model weights, imatrix-calibrated, smaller non-linear 4-bit quant |

| mmproj-QwenPaw-Flash-9B-heretic-BF16.gguf | BF16 | 922 MB | b9821 | 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.

llama.cpp CLI (with image)

./llama-cli \
  -m QwenPaw-Flash-9B-heretic-Q4_K_M-imatrix.gguf \
  --mmproj mmproj-QwenPaw-Flash-9B-heretic-BF16.gguf \
  -c 8192 \
  -ngl 99 \
  --image "path/to/image.jpg" \
  -p "<|im_start|>user\nAnalyze the attached image and plan the next agent action.<|im_end|>\n<|im_start|>assistant\n"

OpenAI-Compatible API Server

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

---

Prompt Format (ChatML)

Supply tool definitions and memory context in the system block for agentic use.

<|im_start|>system
You are QwenPaw, an autonomous agent with access to tools, memory, and planning capabilities.<|im_end|>
<|im_start|>user
Your task 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.
• Primarily designed for the QwenPaw agent stack; strengths are most evident in agentic, tool-calling contexts.
• Vision tensors are kept at BF16 to prevent spatial feature degradation during visual token extraction.
• Imatrix calibration improves perplexity recovery compared to non-imatrix quantization, particularly on low-frequency tokens.
• IQ4_NL produces a smaller file than Q4_K_M and tends to run faster on CPU and ARM devices; imatrix calibration narrows the quality gap between the two formats considerably.