inferwire
/
AI·4 min read

Antirez's h3.c Runs MiniMax-H3 Video Generation on Apple Silicon

h3.c is a native Metal engine that runs MiniMax's open-weights video and audio model on Apple Silicon Macs. Its README documents each speed, memory and quality trade-off with measurements.

TL;DR

  • h3.c runs MiniMax-H3, an open-weights model that generates video with sound, natively on Apple Silicon through Apple's Metal GPU interface [^1] [^2].
  • Its README measures each shortcut, from a four-pass mode to streaming weights from the SSD, which cut tracked transformer storage to 2.0 GiB [^1].

Background

MiniMax released H3 on August 3. It is the company's third-generation video model, after Hailuo 01 and Hailuo 02, and the first it has released with open weights [^2]. H3 takes text, images, video or audio and generates video with stereo sound, according to ComfyUI, which supported it on day one [^2]. Models like this are large: ComfyUI puts the full-precision footprint at 123.6 GB [^2].

What happened

Salvatore Sanfilippo, who publishes on GitHub as antirez, is building h3.c, also called h3-metal, as native MiniMax-H3 inference for Apple Silicon [^3] [^1]. As of August 11, the README says prompt-to-video and audio, first and last frame conditioning and ordered image, video and audio references work end to end, and that current work is Metal performance and memory optimization on M3 Max and M5 Max [^1]. The tool reads the Hugging Face snapshot of the model, needs FFmpeg for media, and can run as an interactive session that keeps the prompt conditioning, the diffusion transformer (DiT) and the video decoder in memory between prompts [^1].

The suggested first run makes 22 frames at 24 fps, about 0.92 seconds of video, at 512 by 512 pixels. It uses 20 denoising passes, computes only 11 of them fresh and extrapolates the rest, and runs 45 of the model's 50 transformer blocks to save time and memory [^1]. A four-pass mode finished denoising in about 3.5 seconds on an M5 Max, against 26.4 seconds for a 29-pass reference, with a full-video SSIM of 0.556 against that reference [^1]. SSIM is a similarity score, where 1 means identical.

For machines with less memory, an SSD streaming mode keeps only two DiT blocks in memory and reads the next one from the SSD while the GPU works, using the original BF16 weights without conversion or quantization [^1]. On an M5 Max, tracked DiT storage fell from about 36.5 GiB to 2.0 GiB at 512 square. A warm 50-block pass took 2.49 seconds instead of 1.35, 84 percent slower, and the output was byte-identical [^1]. The README stresses that the 2.0 GiB figure covers the DiT's tensors only, not total system RAM [^1].

Why it matters

The project shows one way to bring a large open-weights model to Apple hardware: by trading quality and speed in steps that are each measured and documented, rather than by one headline number [^1]. ComfyUI took a different route for consumer GPUs, pruning and quantizing the weights to cut the footprint to 42.5 GB for its smallest variants [^2]. h3.c's streaming mode instead keeps the original BF16 weights and pays for the lower memory with time [^1].

The limits are clear from the README. The hardware it is tuned on is the high end of Apple's range, M3 Max and M5 Max, and the M5 machine it targets has 128 GiB [^1]. It describes H3-Base as a 768p model, caps the canvas at 768 by 1344 pixels, and treats 512 square, sub-second clips as the safe development setting, while the released workflow aims at roughly 4 to 15 second videos [^1]. Its tests check Metal blocks against MLX reference outputs, and the README says pixel-identical videos are not expected because the random-number and execution engines differ [^1].

Practical example

Say you have an M5 Max Mac and want to rough out a short clip of a fox walking through snow. You start with the README's fast preset at 512 square and 22 frames, and turn on the live preview, which shows a frame after every denoising pass in terminals such as Ghostty, iTerm2 or WezTerm [^1]. Once the composition looks right, you rerun the same prompt and seed with all 50 blocks and 50 passes, the slower path the README calls the right reference when a fast mode changes the subject or motion [^1]. If memory runs short, you add SSD streaming and accept the slower passes [^1].

Related gear

We recommend this book because it explains how hardware, memory hierarchies, and low-level C code interact to maximize compute efficiency.

AdvertisementAmazon

Computer Systems: A Programmer's Perspective

Sources

  1. [1]GitHub — h3.c/README.md at 8974cc0 (antirez/h3.c)
  2. [2]ComfyUI — MiniMax H3 Day-0 Support in ComfyUI: Open Weights, Native Audio, and 2K Video
  3. [3]GitHub — antirez (Salvatore Sanfilippo)
  4. [4]Hacker News — H3-metal – Native MiniMax-H3 inference for Apple Silicon