MiniMax H3 Runs Locally Across Four RTX 3090s, Cutting Video Time by 2.84x
An independent tinkerer publishes ComfyUI workflows, custom nodes, and benchmarks that push MiniMax H3 video generation to 3.12x speedup across four RTX 3090s.
- Independent developer alesha-pro published ComfyUI workflows and benchmarks for MiniMax H3 on a 4x RTX 3090 rig.
- Multi-GPU workflow hits 3.12x speedup, cutting a 15-second shot from 19 minutes to 6:41.
- Includes custom nodes, a turbo LoRA loading patch, and a three-command installer.
- Two workflow sets: one for any single 24 GB card, one for multi-GPU boxes.
- Targets MiniMax H3, an open-weights model that generates video with synchronized stereo audio in one pass.
- MIT licensed, with raw benchmark logs from real runs rather than datasheet estimates.
MiniMax H3 toolkit scales local video generation across four RTX 3090s
Independent developer @superalesha has released an MIT-licensed GitHub repository for running MiniMax H3 through ComfyUI, the node-based interface for generative media. Built and benchmarked on four RTX 3090 cards, the toolkit includes 11 workflows, custom nodes, a Turbo LoRA loading patch, and setup scripts.
The published benchmark reduces sampling for a 15-second clip from 19 minutes on one card to 6 minutes 41 seconds on four cards connected through PCIe 3.0 x16 without NVLink. That result gives developers a concrete reference for evaluating multi-GPU H3 inference on consumer hardware.
An installable H3 workbench
The minimax-h3 directory packages the components needed to reproduce the author’s local setup:
| Component | Purpose |
|---|---|
| 11 ComfyUI workflows | Cover single-GPU and multi-GPU generation paths. |
| Single-GPU configurations | Target NVIDIA cards with 24 GB of VRAM. |
| Multi-GPU configurations | Distribute generation work across the available cards. |
| Custom nodes | Provide workflow-specific loading and execution behavior. |
| Turbo LoRA patch | Fixes a known loading problem with the acceleration adapter. |
| Setup scripts | Check dependencies and prepare the local environment. |
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
