MiniMax H3 Shrinks to 8.9 GB so Consumer GPUs Can Run It
A community release packs MiniMax H3's omni-modal video generation into 8.9-21.6 GB GGUF files that run on consumer GPUs through ComfyUI.
- Community release puts MiniMax H3 on consumer GPUs via pruned GGUF quantizations.
- Checkpoints range from 8.9 GB (Q3_K_M, ~12 GB VRAM) up to 21.6 GB (Q8_0 reference).
- Two modes shipped: FL2VA first/last-frame-to-video-audio and Ref2VA omni-reference generation.
- Base MiniMax H3 is a 26B omni-modal model producing 15-second 2K video with native stereo audio.
- Designed for ComfyUI; drop files into models/unet and load with a GGUF UNet loader.
- Released under the MiniMax H3 Community License, permitting commercial use under $20M revenue with attribution.
A community packager has released quantized H3 weights designed for local ComfyUI workflows. The MiniMax-H3-Pruned-GGUF collection combines an already-pruned checkpoint with GGUF quantization, reducing individual files to between 8.9 GB and 21.6 GB.
Running H3 locally removes API charges, cloud queues, and external uploads from the generation loop. The repository had recorded more than 840,000 downloads at the time of publication, although Hugging Face counts file transfers and may include repeat or automated downloads.
Cloud launch, local route
MiniMax introduced H3 as a 26-billion-parameter multimodal system that accepts text, images, video, and audio, then generates clips with native stereo sound. According to the H3 announcement, the model supports clips up to 15 seconds and output resolutions up to 2K.
Initial access came through MiniMax’s API and Hailuo consumer app, with downloadable weights promised later. A 26-billion-parameter checkpoint would require roughly 52 GB for FP16 weights alone, before accounting for inference overhead, which places an uncompressed build beyond most consumer GPUs. The community package starts from the smaller pruned checkpoint and reduces its precision further.
Two modes, five sizes
GGUF is a model file format that supports lower-precision weights and compatible memory-mapped loaders. Lower precision reduces storage and memory use, while higher quantization levels preserve more numerical detail at the cost of larger files.
The repository provides separate checkpoint families for H3’s two main generation modes:
| Mode | Purpose |
|---|---|
FL2VA |
Generates video and audio from first and last frames, prompts, and supporting inputs. |
Ref2VA |
Uses reference images, videos, and audio clips to guide the generated video and soundtrack. |
Each mode is available at five quantization levels:
| Quantization | File size | Repository guidance |
|---|---|---|
Q3_K_M |
8.9 GB | Approximately 12 GB of VRAM with memory-saving settings |
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
