Vibeleaderboard
← All Intel
Intel / repo

MiniMax H3 on a 16GB Mac, 5 days after open weights

Source
github.com
Author
t-go
Date

Context

VPIPE, an app for running AI models locally on Apple Silicon Macs, runs MiniMax's H3 video-and-soundtrack model on as little as 16 GB of memory. Its technical documentation explains that the weights are streamed from storage, so they need not all fit in RAM, and that rereading them adds storage work on every generation step.

That makes local generation possible without making it quick. The README reports roughly thirteen minutes for a five-second, 0.5-megapixel clip with six generation steps on a fanless 16 GB M5 MacBook Air. It discloses that an ice pack under the chassis helped, and that a stock Air will throttle sooner and take longer.

Terms in this piece · Glossary
  • multimodal — A model that works with more than text — reading images, audio, or video, and sometimes generating them too.
  • streaming — Sending a model's response token by token as it is generated, so the reader sees text immediately instead of waiting for the whole answer.
Recommended reads
Comments

Checking sign-in…

Loading comments…