August 10, 2026
Your Mac, but make it dramatic
Antirez/h3.c: MiniMax H3 inference engine for Mac computers
Mac video AI lands, and the comments instantly turned into a rich-kids memory roast
TLDR: antirez has a new Mac tool that can generate AI video and audio locally on Apple Silicon, pushing high-end Macs further into creator territory. The comments quickly split between people joking that the memory demands are absurd and others insisting it already runs surprisingly well with the right setup.
A new project from Redis creator antirez aims to make MiniMax H3 run natively on Apple Silicon Macs, letting people generate short videos and audio straight from their machines instead of leaning on cloud services. On paper, it’s a very nerdy release: build steps, command-line prompts, image references, first-and-last-frame controls, and lots of work squeezing better speed out of high-end Mac chips. But in the comments, the real show wasn’t the code — it was the collective gasp over how much memory this still seems to want.
The biggest mood? Equal parts awe and wallet pain. One user joked that with “only” 96GB of memory they were apparently “a commoner,” which perfectly captured the thread’s main drama: is this a cool Mac breakthrough, or just another toy for people with absurdly expensive machines? Then came the pushback. Another commenter said they’re already running it on a 64GB MacBook Pro through ComfyUI with a smaller model version, basically telling the panic brigade to calm down and tweak their setup.
Meanwhile, the cult of antirez was fully activated. “wow antirez does not sleep” was the vibe of fans watching yet another ambitious side quest appear out of nowhere. There was even a totally random comedy detour comparing Jeff Dean and Salvatore Sanfilippo, because of course every tech thread eventually turns into lore. And one hardware realist tossed in a spicy jab that for this kind of image-and-video AI work, Nvidia-style systems still have an edge — a reminder that even in a Mac victory lap, the platform wars never stay buried for long.
Key Points
- •h3-metal is a native Apple Silicon inference engine for MiniMax-H3 being built in staged vertical slices, with several core generation and conditioning features already working end to end.
- •The current development focus is H3-specific Metal performance and memory optimization on M3 Max and M5 Max hardware.
- •The tool can inspect a local MiniMax-H3 model snapshot and run either as a command-line generator or an Iris-style interactive session.
- •The interactive session keeps BF16 prompt conditioning, DiT, and the video decoder resident in memory and supports persistent first/last-frame anchors or ordered Ref2VA references.
- •The article provides a balanced preset for generating a short 22-frame 24 fps video and explains how options such as --reuse, --layers, --show, and --profile affect speed, memory use, preview behavior, and benchmarking.