Our Hugging Face tracker’s biggest mover this morning was FastVideo/FastVideo-FastH3-Comfy, up 414% in a day — from 10,896 downloads to 56,039. It isn’t a new model. It’s the same FastH3 distillation of MiniMax-H3 we covered when it appeared, repackaged so ComfyUI users can actually run it.
That distinction is the story, and it’s the second time this week we’ve watched it happen.
What’s in the repackage
- Files laid out for ComfyUI’s folder structure:
diffusion_models,text_encodersandvae - Pruned bf16 for people with the VRAM
- int8 and NVFP4-AWQ quantizations for people without
- Based on FastH3 8-Step V2, meaning video generation in eight denoising steps rather than the dozens a base diffusion model needs
- Licensed under the MiniMax-H3 community license — check its terms before commercial use, since it isn’t a standard open licence
- MiniMax-H3 has 137 fine-tuned derivatives on the Hub, which says something about where video work is currently centred
Why the ComfyUI build is the moment that matters
A model’s release is a research event. Its ComfyUI packaging is the adoption event.
ComfyUI is where most people who make things with generative video actually work, because a node graph is how you combine a model with everything else a real project needs: masks, conditioning, upscaling, interpolation, control inputs, batch runs. A model that only runs from a Python script is a model most artists will read about and never use.
We saw the identical curve with YuE2 over the past week: original repo Sept 9, ComfyUI packaging up 142% two days running, GGUF quantization for local runtimes up 31%. Now FastH3 does it with a 414% single-day jump. The derivative builds — Comfy packages, GGUF conversions, quantized variants — are a better adoption signal than the original repo’s stars or likes, and they’re the part of the pipeline that decides whether a model reaches people who make work rather than papers.
The practical read
Eight-step generation plus int8 or NVFP4 weights is aimed squarely at people generating video on one consumer GPU. The tradeoffs are the usual ones: distillation costs you some quality and diversity against the full model, and aggressive quantization costs a bit more. For iteration — trying twenty variations of a shot before committing — that trade is almost always worth taking, and you can run a final pass at higher precision.
If you’re already in ComfyUI, this is a drop-in afternoon experiment. If you’re not, the wrapper is the reason the rest of the field is.