🧠 Qwen 3.5 Abliterated for ComfyUI (MXFP8 & NVFP4)
Welcome! This repository provides ComfyUI-ready, abliterated versions of Qwen 3.5, optimized for local AI workflows, assistants, and multimodal use inside ComfyUI.
📦 Quantized Model Files
| Model | Precision & Notes | Approx Size |
|---|---|---|
| Heretical‑Qwen3.5‑9B‑fp8.safetensors | FP8 quantized ablated Qwen‑3.5 9B | 11.9 GB |
| qwen3.5_9b_abliterated_nvfp4.safetensors | NVFP4 quantized ablated Qwen‑3.5 9B | 8.36 GB |
| Qwen3.5‑4B‑heretic‑fp8.safetensors | FP8 quantized ablated Qwen‑3.5 4B | 5.51 GB |
| qwen3.5_4b_nvfp4.safetensors | NVFP4 quantized Qwen‑3.5 4B | 3.54 GB |
| qwen3.5_4b_claude46opus_abliterated_mxfp8mixedfp8.safetensors | Mixed FP8 ablated (Claude‑4.6+Opus style) | 5.91 GB |
| qwen3.5_4b_claude46opus_abliterated_nvfp4.safetensors | NVFP4 variant of the above | 3.54 GB |
Example of simple workflow using Qwen 3.5 as an image assistant⬇️⬇️⬇️⬇️⬇️
Example of simple workflow plus image captioning⬇️⬇️⬇️⬇️⬇️
🚀 Overview These models are based on Qwen 3.5, a powerful multilingual LLM family designed for reasoning, coding, and general AI tasks.
This release includes:
🔓 Abliterated variants (uncensored / no refusal behavior) ⚡ MXFP8 & NVFP4 quantizations 🧩 Native ComfyUI compatibility
Abliteration removes built-in refusal behavior without retraining, preserving most of the model’s original capabilities while enabling unrestricted outputs. ✨ Features 🧠 Full AI assistant inside ComfyUI 🖼️ Image understanding (multimodal support) ⚡ Fast inference with low VRAM usage 🔌 Plug-and-play with standard ComfyUI nodes
These models can: Answer questions Generate prompts Assist workflows Analyze images directly inside ComfyUI pipelines ⚙️ Quantization Types 🔹 MXFP8 Balanced performance and quality Works on a wide range of GPUs Ideal default option 🔹 NVFP4 Ultra-low precision (4-bit) Massive VRAM reduction and speed gains Best suited for newer NVIDIA architectures Designed for efficient deployment of LLMs with minimal memory footprint 🧩 ComfyUI Integration
✅ These models load directly using:
CLIP Loader (standard node)
No special loaders required.
📦 Installation Update ComfyUI to the latest version Download the model file (MXFP8 or NVFP4)
Place it in:
ComfyUI/models/clip/ Load it using CLIP Loader 🎬 Workflow
👉 A workflow is provided in this repo to help you get started.
It demonstrates:
AI assistant usage Prompt generation Image interpretation
Highly recommended to download and test the workflow to understand full capabilities.
🧠 What is Abliteration?
Abliteration is a technique that:
Removes refusal/alignment layers Keeps original model intelligence intact Does not require retraining
Result: 👉 More freedom 👉 Same core performance
💡 Use Cases 🎥 Prompt generation for video models (LTX, WAN, etc.) 🧩 ComfyUI automation assistant 🖼️ Image captioning & interpretation ✍️ Creative writing / uncensored outputs 🧠 Local AI copilots ⚠️ Notes Abliterated models are community-created, not official releases Use responsibly depending on your application NVFP4 may require newer GPUs for best performance ❤️ Credits Base model: Qwen Team Quantization & conversion: community efforts ComfyUI integration: ongoing community development 🔥 Final Thoughts
If you want a fully local AI assistant embedded directly inside ComfyUI, this setup is one of the most powerful workflows available right now.