FastVideo releases FastMetal-QAD video models for local Apple Silicon generation
AI video model releases
FastVideo releases FastMetal-QAD video models for local Apple Silicon generation
FastMetal-QAD releases three open-source video checkpoints, at 1.3B, 5B, and 14B parameters, built from Wan video models to run natively on Apple Silicon via FastVideo's new MLX runtime.
On an M4 Max with 36 GB unified memory, the 5B model generates an approximately five-second 720p clip in 151.42 seconds with a 9.34 GiB peak, placing 720p generation within 16 GB Macs.
The models use a three-step DMD2 student training process with quantization-aware affine INT8 weights, and ship as both Diffusers safetensors and pre-quantized MLX checkpoints.
Fast mode reduces the 5B model's approximately five-second 720p generation to 47 seconds; cached prompt embeddings skip the cold umT5 encoding step, which takes about 47 seconds for the 5B baseline.
The 1.3B model targets 480p output on 16 GB or larger Macs, while the 14B model supports 480p and 720p but requires 36 GB or more unified memory and measured 601.82 seconds for a 480p clip.
cho.sh