FastH3-Omni on local hardware: 62 seconds on an RTX 5090, 17 minutes on a 3090
Hao AI Lab timed FastH3-Omni on desktop machines: a 5-second 480p clip from one image took 61.9 seconds on an RTX 5090, 192 seconds on a DGX Spark, 16.9 minutes on an RTX 3090 and 19.5 minutes on an M5 Max Mac. Community users report it running on an RTX 4090 and, with a GGUF workflow, on an 8 GB card.
What it does
The post measures one setup on each machine: 480p, one reference image, a 5-second clip. It also explains what makes a request slow and which settings trade quality for speed.
Results and limits
Adding one reference photo made a request 3.3 to 4.4 times slower than plain text-to-video, and the RTX 3090 also spent about 7 minutes loading. The --reference-max-pixels match option cut the M5 Max time from 19.5 to 4.7 minutes but changed the framing and the action. Requests with three images or long clips did not fit in 24 to 32 GB GPU profiles. On X, @OrctonAI reported 480p 15 s in 1m24s and 768p 15 s in 3m48s on an RTX 4090 in ComfyUI.
Worth knowing
@realrebelai shared a low-VRAM GGUF workflow that he runs on an 8 GB card. MiniMax H3: 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.