MiniMax H3 Open Weights Hit ComfyUI: Consumer GPUs Generate Video Locally
Suspicious_Pizza9529 · reddit · 2026-08-04
MiniMax H3 weights are now live on Hugging Face with native ComfyUI support. Thanks to dynamic VRAM optimization and pruning, VRAM requirements have dropped significantly, allowing consumer GPUs like the 4070ti/5070ti to generate full video with audio in under 2 minutes.
Community reports indicate its prompt adherence outperforms Wan 2.2. However, because the original context module is API-locked, whether running the base model locally without it causes major coherence issues remains an open question for testing.
Related event: MiniMax Releases 33B Open-Weight Multimodal Model H3 with Local Deployment(10 posts)→
More from Infra
- Meta releases Muse Spark Open, trained with vLLM-powered RL inference from day one — colinmcnamara · 2026-10-16
- Meta's open Muse Spark needs 493GB of memory — far beyond any consumer laptop — TejasKumar_ · 2026-10-16
- CoreWeave plunges 20% as credit markets price >60% default odds; is this dot-com all over again? — kimmonismus · 2026-10-16
- MiniCPM-V 4.7 35B-A3B hits 17+ tokens/s on iPad with ~3GB memory — techNmak · 2026-10-16
- China rolls out 'token factories' and a 2026-28 token economy action plan — nordicinst · 2026-10-16
- Vercel Sandbox Drives go GA: attach TBs of data in sub-second, fork instantly — cramforce · 2026-10-16