h3.c

// Overview

h3.c (the repository calls the program h3-metal) is a native inference engine for the MiniMax H3 multimodal model on Apple Silicon, written in C with Metal compute kernels. You point it at a local Hugging Face snapshot of the model weights and it generates video and audio from a text prompt on your own machine — no Python stack, no server, one `./h3` binary plus FFmpeg and FFprobe on your PATH.

The project is built as a sequence of working vertical slices: host and model metadata first, then portable Metal block parity, prompt encoding, prompt-to-video/audio, and first/last-frame conditioning. Prompt-to-video/audio, first/last-frame conditioning and ordered Ref2VA image/video/audio references all work end to end, with ongoing work on H3-specific Metal performance and memory optimisation on M3 Max and M5 Max hardware.

Its most distinctive feature is how hard it pushes on the memory/quality/speed trade-off. Flags let you run a subset of the transformer blocks (`--layers`), reuse denoiser velocities instead of computing every pass (`--reuse`), or stream weights from SSD (`--ssd-streaming`), which the README measured as dropping tracked DiT tensor storage from about 36.5 GiB to 2.0 GiB at 512-square resolution on an M5 Max, using the original BF16 checkpoint with no quantization. Without `-p` the same binary opens an interactive session that keeps prompt conditioning, the prepared DiT and the video decoder resident between generations.

// What it does

• Native C plus Metal implementation for Apple Silicon — one binary, no Python runtime
• Prompt-to-video and prompt-to-audio generation, with first-frame and last-frame conditioning anchors
• Ordered Ref2VA references: attach images that the model sees as <Picture 1>, <Picture 2> and refer to by number in the prompt
• Interactive session mode that caches prompt conditioning and decoders so re-rolling a seed skips reloading
• Explicit quality/speed controls — --steps, --reuse and --layers — plus --ssd-streaming for low-memory runs of the unquantized BF16 checkpoint
• In-terminal previews of the evolving frame via the Kitty/Ghostty and iTerm2/WezTerm/Konsole graphics protocols

// Getting started

The examples assume the Hugging Face snapshot of MiniMax H3 is in ./MiniMax-H3 and that FFmpeg and FFprobe are on your PATH.

make -j8
mkdir -p outputs
./h3 --info -d ./MiniMax-H3

./h3 --profile \
  -d ./MiniMax-H3 \
  -p "A red fox walks through fresh snow in a pine forest. Medium tracking shot, natural winter light, realistic fur, soft footsteps and wind." \
  --width 512 --height 512 \
  --frames 22 --steps 20 \
  --layers 45 --reuse 2 \
  --show \
  -o outputs/fox-fast.mp4

./h3 -d ./MiniMax-H3 --width 512 --height 512 --steps 6

h3> !first opening.png
h3> !last ending.png
h3> The camera moves slowly around the subject.

./h3 --profile \
  -d ./MiniMax-H3 \
  -p "A red fox walks through fresh snow in a pine forest." \
  --width 512 --height 512 --frames 22 --steps 20 \
  --layers 50 --reuse 1 --ssd-streaming \
  -o outputs/fox-ssd.mp4

Commands and code are distilled from the project's own documentation — always check the official repo for the latest.

// When to use it

• Generate video and audio from a prompt entirely on a Mac, with the weights and the output staying on your machine
• Iterate on a shot cheaply — four-pass drafts first, then restore layers and denoising passes as the composition settles
• Fit a large multimodal checkpoint onto a machine with limited unified memory by streaming DiT blocks from SSD instead of quantizing
• Read a compact C implementation to understand how a diffusion-transformer video model is actually executed

// How h3.c compares

h3.c alongside other open-source local runtimes tools AI/TLDR tracks, ranked by GitHub stars.