MiniMax H3 video generation on ComfyUI needs 14–15 GB VRAM and over an hour at 1280×704
ayakitodev · reddit · 2026-08-04
A ComfyUI user benchmarked MiniMax H3 reference-to-video generation at low vs. medium resolution with one reference image and the same prompt.
What they found
- A 1280×704 generation was technically possible, but the run would take more than an hour, so they aborted it.
- Their setup used about 14–15.1 GB VRAM and roughly 64–67 GB RAM, with the SSD absorbing much of the load.
- The author tested multiple model variants, including bf16, int8, and cropped int8, but the runtime stayed similarly high.
- The post includes a full sample prompt for a cinematic anime scene, plus the exact ComfyUI workflow details and model filenames.
The takeaway is that MiniMax H3 can run on a 16 GB GPU, but practical throughput is slow and system memory usage is heavy.
More from Infra
- Ibiden’s AI substrate pricing surge sets up a clean earnings asymmetry — tengyanAI · 2026-08-04
- Big Tech’s OpenAI and Anthropic stakes are inflating reported earnings — Kr00ney · 2026-08-04
- Menlo Ventures says AI has entered phase 2, with infrastructure as the real opportunity — mmurph · 2026-08-04
- Podcast says AI CapEx, compute crunch, and debt-financed data centers are squeezing semis — BenBajarin · 2026-08-04
- MiniMax H3 open weights run 32 minutes down to 7.2 minutes on an L40S — ashishsanu · 2026-08-04
- Fluidstack takes its AI infrastructure dinner series to Austin and keeps hiring — MxMnr · 2026-08-04