Abiray Brings MiniMax H3 Video AI to Consumer GPUs With 1M Downloads
A community contributor packaged MiniMax H3, the open-weights omni-modal video and audio model, into GGUF quants that fit on consumer GPUs.
Takeaways
- Hugging Face user Abiray uploaded GGUF quantizations of MiniMax H3, the 33B omni-modal video model.
- Quants cover Q3 through Q8 for both FL2VA and Ref2VA variants, ranging 15.6 GB to 36 GB.
- A pruned companion repo shrinks the UNet to 8.9-21.6 GB for consumer GPUs.
- Bundle includes Qwen3-VL 32B text encoder plus separate video and audio VAEs.
- H3 generates 4-15 second clips at 24 FPS with synchronized 32 kHz stereo audio.
- Released under MiniMax H3 Community License, not a standard permissive license.
A Hugging Face user going by Abiray has released GGUF quantizations of MiniMax H3, the 33B-parameter omni-modal video and audio generator, along with a pruned variant sized for a single consumer GPU. The upload has crossed a million downloads and is what most people running H3 locally through ComfyUI are actually loading.
MiniMax H3 is an open-weights model released earlier this year. It takes text, images, video, and audio as a unified context and generates synchronized video with stereo audio in a single pass, rather than stitching separate models together. At full precision the weights are too large for anything outside a datacenter, which is where Abiray's quantizations come in.
What's in the repository
The Abiray/MiniMax-H3-GGUF repo bundles quantized versions of both H3 base checkpoints plus the supporting encoders and VAEs needed for inference. Two model variants cover different input modes:
- FL2VA — first-and-last-frame mode. Takes zero, one, or two images and fills in the motion between them.
- Ref2VA — omni-reference mode. Accepts up to 9 images, 3 video clips, and 3 audio clips as references, with a 12-file maximum total.