GlennHasABeard runs MiniMax H3 locally on a 12GB RTX 3060
GlennHasABeard reports running MiniMax H3 in ComfyUI on a 12GB RTX 3060 with int8 weights at 20 steps. Each 15-second clip takes about an hour, and his soup demo was upscaled to 4K with Topaz.

TL;DR
- MiniMax H3 ran locally on a 12GB RTX 3060: GlennHasABeard reports int8 weights, 20 steps, and no turbo.
- A 15-second clip takes roughly an hour, with a later timing reply putting typical runtime at about 50 minutes. His soup demo was generated in ComfyUI, then upscaled to 4K with Topaz.
- The shots keep their demands small: a frost-print clip asks for one paw press, while his cat-ears post explains why he avoids whole cats without references.
The official ComfyUI guide targets roughly one-megapixel output for local H3 generation. For the frost shot, GlennHasABeard also handed the job to Claude through MCP, references included.
The 12GB setup
GlennHasABeard gives a refreshingly specific account of the setup:
- GPU: RTX 3060 with 12GB VRAM.
- Weights: int8, which he says makes the model fit his card.
- Sampling: the official model at 20 steps, without turbo.
- Acceleration: an eight-step LoRA is present in the workflow but switched off.
His system screenshot supplies the rest of the hardware configuration:
The 50-minute figure is a reported typical runtime on that machine, rather than a comparison between accelerated and full-step workflows.
One motion per 15-second shot
The soup clip was simple text-to-video, and GlennHasABeard says his notes record getting it in one attempt. Across three examples, the prompts keep the action tightly bounded:
- Soup: “slowly spreads out into a wide soft spiral.” The camera is locked overhead; the cream starts in the center and spreads without stirring.
- Frost: “presses down once into a thin layer of white frost.” One press happens at a set second, then the paw leaves and does not return.
- Yarn: “rolls it slowly to the right, a line of yarn unwinding behind it.” His pleasingly strict limit is “one push, one direction.”
The soup and frost clips are both 15 seconds long, rendered locally in ComfyUI, then taken through Topaz to 4K.
Partial cats
Without a reference, a whole cat could come out any color, he explains. His clips restrict cats to four forms:
- A paw
- Ears
- A tail
- A silhouette
Forty replies about “slop”
Alongside his own renders, GlennHasABeard praised a piece from a weekend art share. He also reported 40 replies to a question about what makes something “slop,” with most answers returning to intent and formulations such as “quantity over quality.”
Claude and MCP
For the frost clip, he says he told Claude what he wanted to make, and it “used the MCP to do the rest with my references.” He also says Claude may have added wording to the prompt.
Workflow downloads and commercial rights
ComfyUI’s native H3 guide provides downloadable workflow JSON and model-download instructions. MiniMax’s repository distinguishes local H3-Base generation from its full 2K workflow, which also uses the Open Platform API.
Commercial use of locally generated outputs requires a MiniMax commercial license, according to ComfyUI’s licensing note. The same documentation says Comfy Cloud generations already include commercial rights.