DownloadAndLoadQWEN2_5_OMNIModel

If you've been living in ComfyUI, this node is where the ground shifts. DownloadAndLoadQWEN2_5_OMNIModel drops Alibaba's Qwen2.5-Omni - a 7B or 3B end-to-end multimodal model that takes text, images, audio, and video in and answers with both text and actual speech - into your graph as a single node. It's not a checkpoint in the diffusion sense. It's a whole assistant, Apache 2.0, that runs 100% on your machine. No API key, no cloud call, nothing. The "download" part of the name isn't marketing: on first run it pulls the weights straight from HuggingFace and caches them in ComfyUI/models/VLM, then loads them into VRAM ready for the pack's run node.

What the knobs actually do

Four inputs, and only two of them really matter for a first workflow.

• model - Qwen/Qwen2.5-Omni-7B or -3B (default 3B). The 3B is the sensible starting point on a mid-range card; the 7B is noticeably better at reasoning and vision but eats a lot more VRAM.
• quantization - none, 4bit, 8bit (default 8bit). This is where the burn happens, because 8bit/4bit loads go through bitsandbytes - which is not in the pack's requirements.txt. If the loader dies on a quantized load with a bitsandbytes ImportError, that's why: pip install bitsandbytes. none sidesteps it entirely but wants roughly 16GB+ of VRAM for the 7B.
• attention - sdpa (default), eager, or flash_attention_2. Leave it on sdpa. Flash attention 2 only helps if you've already built flash-attn, and it is absolutely not worth installing just for this.
• use_audio_output - true by default. This one's sneaky. The Omni architecture is split into a Thinker (the vision/text understanding brain) and a Talker (the speech synth). Set this to false and the code calls disable_talker(), which saves about 2GB of VRAM - but it also forces the run node's return_audio to be false. No audio out, no matter what you ask for downstream. If you're only captioning and describing, turning it off is a free 2GB; if you want the model to actually talk back, leave it on.

The single output is QWEN2_5_OMNI_model, a model object that wires straight into QWEN2_5_OMNI_Run. That's the whole graph: load, run, done.

Installing it

ComfyUI Manager will find it if you search "ComfyUI-Qwen2_5-Omni". Otherwise:

cd ComfyUI/custom_nodes
git clone https://github.com/MakkiShizu/ComfyUI-Qwen2_5-Omni
cd ComfyUI-Qwen2_5-Omni
pip install -r requirements.txt

On the Windows portable build, use python_embeded\python.exe -m pip install -r ... instead. Then restart ComfyUI. Two things about the dependencies: transformers>=4.52.3 is a hard floor - Qwen2_5OmniForConditionalGeneration didn't exist before that version, so an old transformers gives you an ImportError at node load. And the requirements pull in qwen-omni-utils[decord], which is a real dependency stack, not a toy.

Where people get burned

The first run is a download. The 7B is roughly 16GB from HuggingFace, and the loader blocks until snapshot_download finishes - it's not a background task, so ComfyUI looks frozen. Run the first load and go make tea. The 3B is much friendlier.

And one honest note on trust: this is a small, relatively new pack from a solo author, and LLM/VLM nodes are exactly the category that's been weaponized before in this ecosystem. I read nodes.py before recommending it - the only network call is the HuggingFace download of the official Qwen repo. That's the pattern you want: a node that downloads its own weights should do nothing else. Check before you install, always.

Pick the 3B, 8bit, sdpa, audio on if you want speech - and you've got a local assistant in your graph that most of the "cloud AI" space would charge you monthly for.