WanGP v17 runs 15 s of MiniMax H3 at 1080p in 11 GB of VRAM
WanGP, a free local app for open video, image and audio models, rebuilt its memory engine (MMGP v4) in v17. The developer says 15 seconds of MiniMax H3 at 1080p now needs 11 GB of VRAM instead of 25 GB, and 480p fits in 5-6 GB.
What it does
WanGP is a browser-based app for running open models such as Wan, LTX, HunyuanVideo, MiniMax H3, Flux and Qwen Image on consumer GPUs, with LoRA support and many quantized formats. v17 adds the MMGP v4 memory engine, a RAM/VRAM Management tab, and two upsamplers, including an H3 VAE Upsampler that doubles width and height during decoding.
The README says Profile 4 is up to 25% faster with up to 50% less VRAM on large videos.
Results and limits
The VRAM figures are the developer's own measurements, and the coverage says your results depend on the profile, attention backend, settings, system RAM and clip length. It suggests testing at 480p and 720p first.
Several savings are off by default: you choose the MMGP Optimized VRAM Allocator and turn on Smart Memory Pinning yourself.
Worth knowing
Use Sage2/2+ attention, since many optimizations depend on it. For about 20% more VRAM savings at up to 10% slower speed, set Attention Head Split to Medium; leave it off if you need to reproduce an earlier seed exactly. Users in the EU, UK, US and South Korea need to request a license from MiniMax via the H3 License Request Form (platform.minimax.io/h3-license) before use.
Try it yourself
In WanGP: Config > RAM/VRAM Management > choose an MMGP Optimized VRAM Allocator, enable Smart Memory Pinning, restart, then generate [15 s of H3 at 480p] with Profile 4 before trying 1080p.