base_model: coder3101/Qwen3.5-2B-heretic base_model_relation: quantized library_name: gguf license: apache-2.0 language:

• multilingual pipeline_tag: image-text-to-text tags:
• gguf
• llama.cpp
• text-generation
• heretic
• qwen3.5
• mtp
• uncensored
• multimodal
• vision
• abliterated
• conversational
• q4_k_m
• q5_k_m
• q8_0 quantized_by: FadedRedStar

🤖 Qwen3.5-2B-heretic — GGUF

This repository hosts GGUF weights and the associated vision projection matrix for Qwen3.5-2B-heretic, quantized from the source floating-point tensors provided by coder3101/Qwen3.5-2B-heretic.

🔄 Sister Repository: Check out the Imatrix Sister Repository for enhanced precision at lower bit fractions.

[!NOTE] If you plan on using 4-bit or 5-bit variants, consider the imatrix sister repository instead — importance matrix calibration improves logic retention at those bit depths. This repository is best suited if you want the near-lossless Q8_0 build.

ℹ️ Model Profile & Core Features

Qwen3.5-2B is a dense multimodal transformer in the Qwen3.5 lineup by the Qwen Team (Alibaba), balancing compact size with native vision-language understanding, tool calling, and a 262,144-token context window. It targets deployments that need stronger reasoning than the 0.8B model while remaining light enough for single-GPU or high-end edge inference.

The heretic suffix denotes post-processing via Heretic v1.2.0 with the Magnitude-Preserving Orthogonal Ablation (MPOA) method performed by coder3101, which simultaneously targets multiple refusal directions across model layers to achieve a thorough and stable abliteration.

📋 Technical Specifications

🛠️ Heretic Overrides (ARA)

📊 Refusal Bypass Metrics

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

🧮 Numerical & Tensor Formats

📦 Available Model Files

Main model weights

mmproj — vision projector files

mtp — speculative decoding draft files

🎛️ Component Pairing Guide

Download exactly one main weights file:

• Q4_K_M: Balanced 4-bit format suitable for most everyday use.
• Q5_K_M: Higher-fidelity mid-range format recommended as a general default.
• Q8_0: Near-lossless 8-bit format for when memory is not a constraint.

mmproj files (optional): multimodal vision projectors. Pass one via the --mmproj flag in llama.cpp to enable image input.

• BF16 (Recommended): Highest possible image processing accuracy. While older projectors were small, modern vision towers can hover around 1GB. If you are tight on VRAM, it is completely viable to run this on system RAM (CPU) with a minimal performance penalty, saving your precious GPU space for the main model layers.
• Q8_0: Cuts the projector file size and memory footprint in half (~500MB for larger 1GB files). Use this if you prefer to keep the vision tower hosted entirely on your GPU but need to claw back some VRAM to avoid Out-Of-Memory (OOM) crashes.

mtp files (optional): multi-token-prediction draft models for speculative decoding. Pass one via the --draft flag in llama.cpp to speed up generation.

• BF16 (Best Speed): Maximizes draft accuracy. The more accurate the draft model's predictions are, the higher your token acceptance rate, which translates directly into faster text generation. Use this if you have the VRAM headroom.
• Q8_0 (Best VRAM Efficiency): Cuts the draft model size in half. Choose this if loading a heavy BF16 draft model would force you to drop layers of your main model to system RAM, which would heavily tank your total performance.

⚡ Deployment & Execution Commands

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

[!NOTE] Qwen recommends the following sampling configuration for best results:

• Non-thinking mode for text tasks: temperature=1.0, top_p=1.00, top_k=20, min_p=0.0, presence_penalty=2.0, repetition_penalty=1.0
• Non-thinking mode for VL tasks: temperature=0.7, top_p=0.80, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0
• Thinking mode for text tasks: temperature=1.0, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0
• Thinking mode for VL or precise coding (e.g. WebDev) tasks : temperature=0.6, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=0.0, repetition_penalty=1.0

[!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.

[!TIP] Swap the -m filename below for either quantized file depending on your size/quality trade-off preference.

llama.cpp CLI (with image)

./llama-cli \
  -m Qwen3.5-2B-heretic-Q4_K_M.gguf \
  --mmproj mmproj-Qwen3.5-2B-heretic-Q8_0.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"

Speculative Decoding (MTP Acceleration)

./llama-cli \
  -m Qwen3.5-2B-heretic-Q4_K_M.gguf \
  --draft mtp-Qwen3.5-2B-heretic-Q8_0.gguf \
  -c 8192 \
  -ngl 99 \
  -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 Qwen3.5-2B-heretic-Q4_K_M.gguf \
  --mmproj mmproj-Qwen3.5-2B-heretic-Q8_0.gguf \
  -c 16384 \
  -ngl 99 \
  --flash-attn

💬 Chat Templates & Prompt Design (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

⚠️ Safety & Operational Notes

• 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 or Q8_0 depending on the mmproj variant chosen, to preserve spatial feature quality.
• Speculative decoding via the accompanying mtp draft model can meaningfully increase throughput on supported llama.cpp builds.
• For better output quality at this quantization level, consider the imatrix variant in the companion repository.