antirez's h3.c runs MiniMax-H3, a 33B video model, on your Mac
The open 33B model makes video with native audio; a Metal engine, MLX ports, and day-0 ComfyUI now render it locally in ~75s instead of 45 minutes.
Copy markdown
A Redis creator's hand-written Metal engine
antirez (Salvatore Sanfilippo) shipped h3.c, a native MiniMax-H3 inference engine for Apple Silicon. On an M5 Max it renders image and video in about 75 seconds at ~40GB peak memory — well under the 128GB some feared — across 22 to 362 frames at 24fps. You'll need FFmpeg on your PATH.
What the open MiniMax-H3 weight gives you
H3-Base is a 33B dense transformer that generates 4-15s clips at 768p and 24fps with native 32kHz stereo audio, upscalable to 2K through a hosted step (about $0.13/second). It conditions on text, image, video, and audio, with first- or last-frame anchoring and up to 12 reference files.
Pure-MLX ports for Apple Silicon
minimax-h3-mlx is a pure-MLX port validated against the diffusers reference, with an AdaLN precompute that drops ~13B at inference. Simon Willison ran a 15-second clip in ~45 minutes on an M5 Max (115GB of weights) before these faster engines landed.
Day-0 ComfyUI, plus GGUF for 64GB Macs
Comfy-Org shipped native H3 nodes at launch, and quantized GGUF lets a 64GB M5 Pro run it — roughly an hour for a 9-second clip at modest settings. First/last-frame and reference-guided workflows are ready to wire up if you already live in ComfyUI.
Read the license before you ship
The Community License carves out the US, EU, UK, and South Korea without separate authorization, and commercial products earning over $20M a year need written approval. The weights are open, but this is not an unconditional green light.