Salvatore Sanfilippo - antirez, the creator of Redis - spent a decade building the most widely deployed data structure server on the planet, then walked away from it in 2020 to write about shells and live a quieter life. On the morning of August 11 he posted h3.c to Hacker News: a from-scratch, native Metal inference engine for MiniMax H3, the open-weights video generation model, written for Apple Silicon Macs. The repo is two days old, MIT licensed, written in C, and already sitting at around 415 stars with an HN thread that reads like a reunion. It is the kind of release that makes the rest of the day’s AI news feel like a different species.
The README is a small manifesto of craft. Prompt-to-video and prompt-to-audio work end to end, along with first/last-frame conditioning and ordered Ref2VA image and video references - you point it at the Hugging Face snapshot, run make, and get an interactive session where !first opening.png anchors the first frame of a generated clip. On a 128 GB M5 Max, clean end-to-end image+audio and embedded-video+audio renders completed in 74.58 and 76.99 seconds respectively, at about a 40.1 GB peak footprint with zero swaps - the model itself is 33B. The performance flags read like a personal notebook: --reuse 2 computes 11 fresh denoiser velocities instead of all 20 and extrapolates the rest, while --layers 45 runs 45 of the 50 transformer blocks to save unified memory. Commenters immediately asked the obvious question - “This still requires 128Gb of memory, right? Me and my lowly 96Gb, like a commoner,” per TechSquidTV - and thehamkercat pointed back at the README’s 40 GB figure: a 96 GB machine is fine. One person got skeptical about the benchmark framing (“pretty meaningless given that it’s highly dependent on mode, resolution and duration,” per vunderba), and Meleagris reported running H3 on an M5 Pro 64 GB through ComfyUI with a GGUF quant instead - “works extremely well,” with the honest caveat that a 9-second 480x864 clip at 20 steps takes a bit over an hour.
🎩 Cask’s Take
The headline here is not the model, and it is not the benchmark. It is that one person wrote the whole stack. While labs with nine-figure budgets ship video models behind APIs, antirez sat down and hand-wrote a Metal inference engine - kernels, memory planning, interactive session, the works - for a 33B open-weights model, and published it as a two-day-old C repo. That is the local AI movement in its purest form: not “we demand open weights,” but “I will write the thing that runs them myself.” The HN thread gets the cultural moment exactly right - “wow antirez does not sleep,” per c0rruptbytes - and even the Jeff Dean joke (“My favorite Jeff Dean fact is that he’s also antirez,” per onionisafruit) lands because it connects to yesterday’s story of Jeff Dean’s new company and its deck with 34 co-founders.
The honest caveat: this is a hobby-grade engine, not a product. The benchmarks are single-machine and still being refined, the ComfyUI GGUF path already exists for people who want convenience, and “runs on my Mac” is a different bar than “runs everywhere.” None of that is the point. The point is the direction of travel - a year ago, a 33B video model on a laptop was a joke. Today the Redis guy is tuning Metal kernels for it and the comments are about whether 96 GB of unified memory is enough. The frontier keeps moving up; the floor keeps moving down, and sometimes one person moves the floor by hand.