MiniMax H3 video model runs natively on Mac via Metal, 4x faster than PyTorch
DaLyon92x · reddit · 2026-10-24
A developer released MacH3 Metal, a ComfyUI node pack that wraps antirez's C+Metal engine h3.c to run the MiniMax H3 video model natively on Apple silicon, bypassing the slow PyTorch route.
- Features: text-to-video, first/first+last frame, up to 3 reference subjects, and a "Continue" mode that chains takes from the last 22 frames and 0.9s of audio, with a Shot node to join them
- Speed: 6-step Turbo (larryvrh's Turbo LoRA) and 12+ step Quality modes; on a 48 GB M5 Pro, a 3s 544x960 clip takes 3:25 in Turbo — roughly 4x faster per step than PyTorch+GGUF (45-50s vs 188s at 768p)
- Memory: 18 GB of int8 weights are memory-mapped from SSD; ComfyUI peaks at just 7.7-9.1 GB
- Limits: Mac/Metal only, runs outside the latent graph (no Comfy samplers/LoRA nodes); should work on 24 GB+ M1 or newer
MIT-licensed, workflows on Civitai.
More from Embodied
- Enact emerges from stealth with $30M pre-seed to teach robots force and touch — ylecun · 2026-10-24
- Audi tests wheeled humanoids from Agile Robots for factory logistics, with NVIDIA and Dassault support — CyberRobooo · 2026-10-24
- Microduck: a 25cm, 800g desk robot packing 15 motors, LiDAR and dual IMUs — Olivier__OG · 2026-10-24
- Home-trained Microduck robot walks down steps and hills in one gait, trained on an RTX 5090 — alexcovo_eth · 2026-10-24
- Autonomous Lamp: open source companion robot ships at $969 with BYO-AI — dee_hw · 2026-10-24
- Claude Code wrote a smartwatch firmware from scratch in three days — the full workflow — cyberdinn · 2026-10-24