MiniMax H3 local setup and FLUX.3 vs Wan 3.0

MattVidPro tests MiniMax H3 on his own GPU and compares its video output with FLUX.3 and Wan 3.0. He describes H3 as an open-weight model with native audio, 720p output and generations up to 15 seconds. His examples include spoken comedy scenes and a construction timelapse. He praises the audio but points out disappearing objects, inconsistent physics, altered shapes and scratchy sound in some clips.

The RTX 5090 tests give a concrete sense of the wait involved. A five-second clip just below 720p at 24 FPS took roughly three and a half minutes. Longer, 15-second generations reached more than 30 minutes; one 1024 by 768 example took about 25 minutes. Community reports of smaller GPUs are separate from his own tests, with lower resolution and quality tradeoffs.

For local AI setup, he discusses Maestro with a WanGP backend through Pinokio, which requires Nvidia hardware and CUDA. He also cites a DiffSynth Studio route using NF4 weights on Apple M-series hardware with a reported 8 GB requirement, while noting quality loss and disputed community claims. His own installation uses Codex to configure headless ComfyUI and queue generations.

He prefers FLUX.3 for the shared timelapse prompt. At the time described, its open weights were still pending, while Wan 3.0 was an API-only public beta. He also explains that the Heretic encoder needs the H3 weights and raises concerns about likeness and voice imitation.