**h3.c** (internally *h3-metal*) is a from-scratch **C inference engine that runs MiniMax-H3 video and audio generation natively on Apple Silicon**, built by antirez — the author of Redis, and more recently of `ds4` and `h3`-adjacent local inference work. It is barely three days old at time of writing and already past 1,200 stars, following the same pattern as the author's earlier single-purpose engines: no Python, no framework, just Metal and a `make -j8`. The project is being built as a **sequence of working vertical slices** rather than a monolithic port, and the README is unusually candid about which slices are finished. Prompt-to-video and prompt-to-audio, first/last-frame conditioning, and ordered Ref2VA image/video/audio references all work end to end. The current work is H3-specific Metal performance and memory optimization on M3 Max and M5 Max. What makes it interesting beyond the port itself is the **explicit speed/quality control surface**, each dial documented with measured tradeoffs rather than vibes. `--steps` names actual denoising passes; `--reuse` extrapolates skipped transitions (20 steps becomes 20, 11, or 8 fresh DiT evaluations); `--layers` thins the 50-block transformer; `--core-reuse` keeps patch and output heads fresh while running the expensive core less often; `--token-reduction` pairs horizontal video tokens in middle blocks. A validated **four-pass** render measured **0.556 full-video SSIM** against a 29-pass reference and took roughly **3.5 seconds on M5 Max versus 26.4 seconds** for that reference. The docs also record which combinations *failed* — stacking token reduction onto `--layers 40` with `--reuse 3` produced color ringing and ghosted limbs — which is rarer and more useful than a table of wins. The standout memory feature is **`--ssd-streaming`**: it keeps two DiT blocks resident and reads the next block from SSD while the GPU works, using the original BF16 checkpoint with no conversion or quantization. Tracked DiT storage fell from about **36.5 GiB to 2.0 GiB** at 512 square, at a cost of 84% slower warm forwards at that size and 26% slower at 864x480 — and the outputs were **byte-identical** in both checks. There is also an Iris-style interactive session that keeps prompt conditioning, the prepared DiT, and the video decoder resident so reruns with new seeds skip reloading, plus in-terminal preview via Kitty/Ghostty and iTerm2/WezTerm/Konsole graphics protocols. The constraints are real. This is **Apple Silicon only**, with tuning validated on M3 Max and M5 Max — not a portable engine. Width and height must be multiples of 32 with the product capped at 768×1344, frames align upward to `5 + 17n`, and native 128-square is unsupported because a 4×4 token grid never recovered a recognizable subject. Pixel-exact parity with the reference MLX implementation is **not expected**, since the RNG and execution engines differ; the project targets agreeing content and motion instead. **MIT licensed**, and it expects a local Hugging Face MiniMax-H3 snapshot plus FFmpeg on `PATH`.