H3-metal is an inference tool for MiniMax-H3 targeting Apple Silicon. The developer is building functionality as a series of practical vertical slices, and has currently completed the ensuring of compatibility for model metadata and Metal blocks. Prompt-to-video and prompt-to-audio generation, as well as first- and last-frame conditioning, are now operational end-to-end. This tool assumes the use of Hugging Face snapshots and FFmpeg.
In the runtime, the tool is built using make -j8, and device information can be verified via ./h3 --info. If the -p option is not provided, an Iris-style interactive session starts. In this mode, BF16 prompt conditioning and the decoder are held in memory, eliminating the need for reloading when changing seeds. State management and output saving are possible via commands such as !status and !save.
Regarding conditioning features, the !first and !last commands allow for fixing the start and end frames of a video. Additionally, !ref-image allows reference images to be added in sequence and passed to the model as <Picture 1>. However, Ref2VA reference and first/last-frame conditioning cannot be used simultaneously.
In terms of performance, Metal optimization for M3 Max and M5 Max is in progress. Verified presets generate videos of 22 frames at 24fps, with intermediate frames displayed in real-time. Iteration with low step counts, such as --steps 4, is also possible, allowing for the adjustment of the balance between detail and motion.
Sources: H3-metal – Native MiniMax-H3 inference for Apple Silicon (HN 440pt, 99 comments) (HN Search (backfill), 2026-08-11)