This fixed model list is so awful it's made the entire repo useless

[Parogar]

I'm trying to get it to load an abliterated model but it's just not working. Even naming it custom_models.json doesn't help. Honestly, it's like the creators of this repo are pro censorship

[Woisek]

I also wonder how this works or even better, why it has no possibility to choose a model from the system. 🤷‍♂️

[Parogar]

I also wonder how this works or even better, why it has no possibility to choose a model from the system. 🤷‍♂️

it's because they want us to play by THEIR rules and THEIR censorship

[HWiese1980]

Let's make a fork that allows arbitrary models. Open Source for the win!

[HWiese1980]

I've created a Pull Request with a corresponding change. #169

[Angelthree95]

I've created a Pull Request with a corresponding change. #169

If I understand the code correctly it will enable use of a model with a configured path in custom models json "LLM/GGUF/Qwen3-VL-8B-Thinking-abliterated-Q6_K.gguf" to be loaded from "/ComfyUI/models/LLM/GGUF/Qwen3-VL-8B-Thinking-abliterated-Q6_K.gguf"?
Maybe it's a good idea to document a little bit on how to configure it with locally stored models?

[HWiese1980]

Yes, definitely a good idea. I'm gonna write some docs for it. 👍

[Deaquay]

I just edit the

hf_models.json like this:

{
  "hf_vl_models": {
    "Qwen3.5-9B-ultra-heretic": {
      "repo_id": "Qwen3.5-9B-ultra-heretic",
      "default": true,
      "quantized": true,
      "vram_requirement": {
        "full": 12.0,
        "8bit": 7.0,
        "4bit": 4.5
      }
    },
    "Qwen3.5-9B-ultra-heretic-fp8": {
      "repo_id": "Qwen3.5-9B-ultra-heretic-fp8",
      "default": true,
      "quantized": true,
      "vram_requirement": {
        "8bit": 7.0,
        "4bit": 4.5
      }
    },
    "Qwen3-VL-4B-Instruct-Abliterated": {
      "repo_id": "huihui-ai/Huihui-Qwen3-VL-4B-Instruct-abliterated",
      "default": false,
      "quantized": true,
      "vram_requirement": {
        "full": 6.0,
        "8bit": 3.5,
        "4bit": 2.0
      }
    },
etc...

Amd the gguf_models.json with this:

  "qwenVL_model": {
    "Qwen3.5-9B-ultra-heretic-GGUF": {
      "author": "llmfan46",
      "repo_name": "Qwen3.5-9B-ultra-heretic-GGUF",
      "repo_id": "llmfan46/Qwen3.5-9B-ultra-heretic-GGUF",
      "mmproj_file": "Qwen3.5-9B-mmproj-F32.gguf",
      "model_files": [
        "Qwen3.5-9B-ultra-heretic-Q4_K_M.gguf",
        "Qwen3.5-9B-ultra-heretic-Q5_K_M.gguf",
        "Qwen3.5-9B-ultra-heretic-Q6_K.gguf",
        "Qwen3.5-9B-ultra-heretic-Q8_0.gguf",
        "Qwen3.5-9B-ultra-heretic-BF16.gguf"
      ],
      "defaults": {
        "context_length": 262144,
        "image_max_tokens": 4096,
        "n_batch": 512,
        "gpu_layers": -1,
        "top_k": 20,
        "pool_size": 4194304
      }
    },

I have them saved in a separate file and just copy paste when they update the nodes.

The models then go like such for gguf: LLM/GGUF/llmfan46/Qwen3.5-9B-ultra-heretic-GGUF/
And LLM/Qwen-VL/Qwen3.5-9B-ultra-heretic/ for hf_models

But you'll have to patch AILab_QwenVL.py and _GGUF.py to stop the models from reasoning. 3.5 REAALLY likes to think

[Kozatiju]

Maybe it could be usefull for someone :

I ran into into another issue on loading custom gguf models and here how I fixed it :
In the "gguf_models.json", beside modifying the base dir to LLM/GGUF instead of llm/GGUF as @HWiese1980 pointed it at the beggining of the file, another mistake was left on line 340 :
base_dir = _resolve_base_dir(GGUF_VL_CATALOG.get("base_dir") or "llm/GGUF")
of course, it had to be changed to :
base_dir = _resolve_base_dir(GGUF_VL_CATALOG.get("base_dir") or "LLM/GGUF")

[raffetazarius]

But you'll have to patch AILab_QwenVL.py and _GGUF.py to stop the models from reasoning. 3.5 REAALLY likes to think

Would you mind explaining how to do this please?

[Deaquay]

But you'll have to patch AILab_QwenVL.py and _GGUF.py to stop the models from reasoning. 3.5 REAALLY likes to think

Would you mind explaining how to do this please?

You can just drop the file(s) to gemini or smth and tell it that if model name starts with "Qwen3.5-" it needs to add enable_thinking = false to chat template.

Should looke something like this:

        is_qwen35 = model_name.startswith("Qwen3.5-")
        chat_kwargs = {}
        if is_qwen35:
            chat_kwargs["enable_thinking"] = False
            print("[QwenVL] Qwen3.5 detected: Disabling thinking in chat template.")

        # Optimize chat template for memory efficiency
        chat = self.processor.apply_chat_template(
            conversation,
            tokenize=False,
            add_generation_prompt=True,
            **chat_kwargs
        )

Also need to change def generate to pass model name as well

    def generate(
        self,
        prompt_text,
        image,
        video,
        frame_count,
        max_tokens,
        temperature,
        top_p,
        num_beams,
        repetition_penalty,
        model_name="",
    ):

Send me a message and I can give you already patched files. Easier that way.

[Nanfeng-tech]

But you'll have to patch AILab_QwenVL.py and _GGUF.py to stop the models from reasoning. 3.5 REAALLY likes to think

Would you mind explaining how to do this please?

You can just drop the file(s) to gemini or smth and tell it that if model name starts with "Qwen3.5-" it needs to add enable_thinking = false to chat template.

Should looke something like this:

        is_qwen35 = model_name.startswith("Qwen3.5-")
        chat_kwargs = {}
        if is_qwen35:
            chat_kwargs["enable_thinking"] = False
            print("[QwenVL] Qwen3.5 detected: Disabling thinking in chat template.")

        # Optimize chat template for memory efficiency
        chat = self.processor.apply_chat_template(
            conversation,
            tokenize=False,
            add_generation_prompt=True,
            **chat_kwargs
        )

Also need to change def generate to pass model name as well

    def generate(
        self,
        prompt_text,
        image,
        video,
        frame_count,
        max_tokens,
        temperature,
        top_p,
        num_beams,
        repetition_penalty,
        model_name="",
    ):

Send me a message and I can give you already patched files. Easier that way.

Could you please post the patched files here? I really need them, thank you.

[Deaquay]

Let me just debug them to make sure there's no errors when my gpu is free.

But you'll have to patch AILab_QwenVL.py and _GGUF.py to stop the models from reasoning. 3.5 REAALLY likes to think

Would you mind explaining how to do this please?

You can just drop the file(s) to gemini or smth and tell it that if model name starts with "Qwen3.5-" it needs to add enable_thinking = false to chat template.
Should looke something like this:

        is_qwen35 = model_name.startswith("Qwen3.5-")
        chat_kwargs = {}
        if is_qwen35:
            chat_kwargs["enable_thinking"] = False
            print("[QwenVL] Qwen3.5 detected: Disabling thinking in chat template.")

        # Optimize chat template for memory efficiency
        chat = self.processor.apply_chat_template(
            conversation,
            tokenize=False,
            add_generation_prompt=True,
            **chat_kwargs
        )

Also need to change def generate to pass model name as well

    def generate(
        self,
        prompt_text,
        image,
        video,
        frame_count,
        max_tokens,
        temperature,
        top_p,
        num_beams,
        repetition_penalty,
        model_name="",
    ):

Send me a message and I can give you already patched files. Easier that way.

Could you please post the patched files here? I really need them, thank you.

Let me just debug them to make sure there's no errors when my gpu is free.
I'll make a fork of the Qwen-VL-Mod fork and add 3.5 support since he has closed issues so you can't communicate with him. And post the link here, then you can just download the files you need. I'll write what files to grab in the readme if you don't want to clone the whole repo.

[Deaquay]

Patched version which accepts any 3.5 model as long as you follow instructions:

ComfyUI-Qwen3.5-Uncensored

Just create an issue in the repo with a list of models you want added or if something is broken.

@Nanfeng-tech @raffetazarius

[raffetazarius]

@Deaquay - you're a LEGEND! Thank you! 🙌

[Deaquay]

Model list is now dynamic. Reads GGUF and Qwen-VL folders for models, and adds them without any setup to the nodes.