HY-Motion Load LLM (GGUF)

If you're not sitting on 24GB of VRAM, this is probably the LLM loader you actually want. HY-Motion's text encoder is Qwen3-8B - a huge model for the job of understanding one motion sentence - and the GGUF loader is the pack's answer to that. Instead of a full transformers model directory hogging 16GB, you point this node at a single quantized .gguf file and it loads the same encoder at a fraction of the memory. Q4_K_M runs at roughly 5GB; Q8_0 is near-lossless at ~8GB. For a motion skeleton, the quantization quality debate barely matters - grab Q4_K_M and move on.

It's a drop-in replacement for the plain Load LLM node: same llm output, same place in the graph, feeds HY-Motion Encode Text exactly the same way. That's the entire point of the pack's GGUF support, which the community was loudly asking for within hours of launch - nobody wanted the full Qwen just to animate five seconds of walking.

Inputs

• gguf_file - an enum filled by scanning your GGUF folders. Files live in ComfyUI/models/HY-Motion/ckpts/GGUF/ (it also checks models/llm/GGUF). The dropdown shows "(select file)" until you actually drop one in and refresh.
• device_strategy - gpu, cpu, or balanced. gpu is the default and the obvious choice; cpu runs the encoder in system RAM so the GPU stays free for the diffusion network, at the cost of slow encoding; balanced splits the difference.

What to download and where

GGUF files are not bundled and the README is explicit that you fetch them manually from huggingface.co/Qwen/Qwen3-8B-GGUF. The pack's recommended files, by size:

• Qwen3-8B-Q4_K_M.gguf (5.03 GB) - the recommended balance of quality and size
• Qwen3-8B-Q5_K_M.gguf (5.85 GB) - a step up
• Qwen3-8B-Q6_K.gguf (6.73 GB) - near-original quality
• Qwen3-8B-Q8_0.gguf (~8 GB) - practically lossless, but then you're back near full-encoder memory

Put the file in ComfyUI/models/HY-Motion/ckpts/GGUF/ and restart ComfyUI so the enum sees it.

The catches

Two things to know before you wire it up. First, it needs recent transformers (>= 4.40) because it uses transformers' native GGUF support rather than llama.cpp - the pack's requirements install this for you, but if you've pinned an old transformers for another node, this one will complain. Second, there's no tokenizer inside a .gguf by itself: the loader tries the GGUF's directory first, then hunts through your other Qwen3 model folders for one, which is why having some Qwen3 tokenizer on disk makes it much more forgiving. When it works you'll see the node doing an aggressive memory cleanup pass before loading - it's genuinely trying not to OOM you.

It's the loader most people should default to. Same encoder, a quarter of the memory, and the only real downside is one manual download you'd have to do anyway.