Hao AI Lab FastH3 is an open-weight local video generator that ports MiniMax H3 to NVIDIA DGX Spark and Apple Silicon via MLX for creators and engineers who want fast video generation without a cloud bill. Released in early September 2026, it reaches up to 8x speedup over the base MiniMax H3 model on local hardware.
Core Features
- MLX and CUDA builds that run on Apple Silicon and NVIDIA DGX Spark from a single model definition.
- Quantized INT8, INT6, and INT4 weights that shrink memory use so longer clips fit on consumer devices.
- Phased weight loading that streams model files instead of holding them all in RAM at once.
- A FastVideo Cookbook with step-by-step local deployment recipes.
- Demonstrated output of a 15-second 1344x768 clip on a single Spark device.
Use Cases
- Makers who iterate on short video concepts offline and want zero per-generation API cost.
- Researchers benchmarking MiniMax H3 speedups on accessible hardware.
- Studios with privacy or export rules that forbid sending footage to external servers.
Pricing
FastH3 is open-weight and free to run; the only cost is your own Mac or DGX Spark and electricity. There is no subscription or per-second fee, unlike hosted video APIs.
Our Take
Best for tinkerers with an M-series Mac or a DGX Spark who want MiniMax H3 quality at home. The trade-off is that 8x speedup still means real wait times for long clips, and you manage your own environment. Browse more AI video tools for cloud alternatives.
