MiniMax H3 W4A8 experimental models

MiniMax-H3-experimental is a Hugging Face repository by Kijai containing experimental compressed checkpoints of the MiniMax H3 video generation model, packaged for ComfyUI. The main files use a W4A8 format (4-bit weights, int8-convrot activations) and require ComfyUI 0.31.0. The repository also lists a work-in-progress W6A8 variant (6-bit weights), with a related comfy-kitchen pull request referenced.

Additional files include int8-convrot quantized video VAEs, which the author says speed up VAE decoding (about 2x for the main VAE with ComfyUI 0.36.0, about 1.3x for a pruned "light" VAE). The repo also holds converted LoRAs: PDMD 2NFE/4NFE and DMAD distillation LoRAs, a rank-reduced extraction of the Singularity v2 finetune (the author notes it captures perhaps 40-50% of the full finetune), an experimental reference LoRA, and a 2K refiner component only.

According to a YouTube review, the W4A8 models are about 12 GB with a roughly 3 GB int8 VAE, and with Spectrum and sage attention can run on about 8 GB of VRAM. The files are labelled experimental by the author; license and model card details were not provided on the page.