Open-source project h3-metal has implemented native MiniMax-H3 inference for Apple Silicon, with prompt-to-video and audio generation, first- and last-frame conditioning, and ordered image, video and audio references working end to end. Development is now focused on Metal performance and memory optimization for M3 Max and M5 Max systems.
The command-line engine uses the original BF16 model checkpoint and expects a local MiniMax-H3 snapshot plus FFmpeg and FFprobe. An information command checks model structure and reports the selected Metal device without loading all weights. Without a prompt flag, the same binary starts an interactive session that retains prompt conditioning, the prepared diffusion transformer and video decoder between requests.
Users can set persistent opening and closing images or add ordered general references that appear to the model as numbered pictures. The two conditioning modes cannot be mixed. Generated files are written to a session directory, and commands can change seeds, inspect status, display output and save a selected video.
A documented balanced preset generates 22 frames at 24 frames per second, about 0.92 seconds of video, using 20 denoising passes and 45 of 50 transformer blocks. Reusing selected denoiser results reduces fresh evaluations. An optional terminal preview supports several graphical protocols but keeps an extra preview decoder resident and adds roughly 10 GiB of temporary model storage.
For rapid iteration, four denoising passes produced a 512-pixel-square fox sample in about 3.5 seconds on an M5 Max, compared with 26.4 seconds for a 29-pass reference. Full-video structural similarity was 0.556 for that sample and 0.547 in a separate surfer test. Those measurements describe selected validation prompts, not general quality guarantees.
An SSD-streaming mode lowers resident transformer storage by keeping two blocks in memory and loading the next from storage while the GPU runs. On M5 Max, tracked transformer tensor storage fell from about 36.5 GiB to 2.0 GiB at 512 square and 2.1 GiB at 864 by 480. Warm forward passes were slower, making the option an explicit memory-for-speed exchange. Total system memory remains higher because operating-system, media, encoder and decoder needs are separate.
The engine also offers layer reduction, velocity reuse, token reduction and smaller internal canvases. The documentation warns that aggressive combinations can alter composition, motion and anatomy or create visual artifacts. h3-metal is therefore both a working inference implementation and a tuning environment: users can choose quicker previews or more expensive reference paths while comparing each change against the model's fuller output.
Resolution limits require width and height to be multiples of 32, with a maximum pixel product of 768 by 1,344. Those mechanical limits do not promise useful quality at every small canvas, since H3-Base was designed as a 768p model.
