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:

Each mode is available at five quantization levels: