Thirteen billion parameters, computed once
A day after MiniMax published its video-and-sound model for anyone to download, a port had it generating clips on a MacBook, mostly by never loading the forty-odd gigabytes that sit idle while it runs.
Simon Willison asked for a rainbow-coloured skunk leaping over a mossy log in a supermarket, waited about three quarters of an hour, and got back a video with a soundtrack, made entirely on the laptop on his desk. The model that made it, MiniMax H3, had gone public the day before as more than a hundred gigabytes of files, built for the sort of machine you rent by the hour.
"no animal, no log, just a textured field" — the repo's README, on what three-bit compression does to the picture
Most of the shrinking came from not loading things. Around 13 billion of the model's parameters exist only to produce a set of adjustment numbers that depend on nothing except how far along the generation has got; compute them once at the start, keep the 745-megabyte result, and 26 gigabytes of weights become surplus. That precomputation takes 0.7 seconds. Another 16 gigabytes go because the part that reads your prompt is a borrowed language model whose last fourteen layers this pipeline never consults, so loading simply stops at layer fifty. Squeezing what remains down to four bits per number leaves about 11.5 gigabytes to hold in memory, from a starting point near 144.
It is not fast. Willison's clip took roughly 45 minutes, and the project's own timings, measured on a desktop Mac with half a terabyte of memory, run into hours at the default quality setting. His audio came back as speech-like babble, because he never told it what to say.
What lingers is the accounting. A released model's file size measures what the lab shipped, not what has to sit in memory while it runs, and finding the gap between the two is nobody's job until someone with a laptop and a free weekend goes looking. That work has been routine on text models for a couple of years. Doing it to a video-and-sound model one day after the weights appeared is where local generation goes next.
The lenses
The facts
How this connects
Tap a node to open it