Salvatore Sanfilippo, the programmer known for building Redis, has published h3-metal, an engine that runs MiniMax’s H3 video generation model directly through Apple’s own graphics framework instead of a rented cloud GPU. The project lives on GitHub under his handle, antirez, and its premise is unusual for this category of software: it assumes the person running it owns the hardware already.
Generating video with a frontier model normally means provisioning a data-center GPU, installing a CUDA toolchain, and paying by the hour, because the underlying transformer needs more memory and throughput than a personal machine typically offers. Sanfilippo’s engine talks to Apple’s Metal compiler instead, so the same generation pipeline that usually lives in a cloud rack can run on a Mac Studio or a high-end MacBook Pro sitting on a desk.
The engine handles three things end to end, according to the repository: turning a text prompt into a short video with matching audio, anchoring a clip to a chosen first and last frame, and building a scene from an ordered set of references drawn from images, silent clips, or clips that keep their own soundtrack. None of the three is a proof-of-concept demo. The remaining work described in the README is narrower: tuning speed and memory specifically for two Apple chips, the M3 Max and the newer M5 Max.
That tuning is where the project earns its credibility. On a Mac configured with 128 gigabytes of memory, a full render combining an image reference with audio finished in roughly 75 seconds and used about 40 gigabytes of memory at its peak, without spilling to disk. Getting there took granular engineering: converting internal matrix operations to 8-bit integers, dropping the transformer blocks that contribute least to the output, and skipping redundant computation across denoising passes. Sanfilippo documents each change with before-and-after timing and a visual comparison, a level of rigor closer to a systems paper than a typical open-source README.
The significance is not raw speed. It is that one developer, working without a company or a cluster behind him, moved a capability that used to require a cloud contract onto a machine an individual can own outright. A startup wanting to prototype with video generation today still defaults to spinning up GPU instances and managing an unfamiliar CUDA environment while the meter runs. Sanfilippo’s tutorial instead assumes a downloaded model snapshot, a copy of FFmpeg, and a Mac.
The honest caveat sits inside the same numbers that make the project impressive. The model’s weights alone occupy 37 gigabytes on disk, and a full-resolution render pushes peak memory past 40 gigabytes even with every optimization switched on. That fits comfortably on a 128 gigabyte Mac Studio. It does not fit on the 16 or 24 gigabytes most consumer laptops ship with, and Apple’s unified memory, more flexible than a discrete GPU’s dedicated VRAM, is still not unlimited. The constraint on who can run this locally was never really the Metal code. It is how much memory a buyer can afford to configure into the machine.
For any team weighing whether video generation belongs in an internal tool instead of a vendor API, h3-metal is working proof that the compute problem is solvable on hardware a company already owns. The decision that follows is not which GPU cloud to rent. It is whether the Mac configuration on hand clears the memory bar the model actually needs, and if not, which tier to buy.
Sourced from the antirez/h3.c repository on GitHub, maintained by Salvatore Sanfilippo, the creator of Redis.
