ZImageApp v2.1.0 · Apple Silicon · macOS 15+
FP16 Dense acceleration on Mac with MLX
Explore optional FP16 Dense computation for supported image models in ZImageApp on Apple Silicon.
Free download v2.1.0 · Apple SiliconDeveloper ID signed and Apple notarized. Models download separately.
FP16 Dense on your Mac
ZImageApp offers an optional FP16 Dense path for supported model layers. It uses half-precision dense computation where supported by the native MLX pipeline, giving you another performance option alongside the default quantized path.
Dense coverage differs by model, including FLUX 3, Krea 2 and Iris. Expanding quantized weights can require more unified memory. Speed depends on your Mac, the model, image size and available memory; the switch does not imply a fixed speedup or identical coverage across models.
How to get started
- Load a compatible model in the native generation panel.
- Enable its FP16 Dense option when available and allow sufficient unified memory.
- Compare generation time and output with the default mode using the same prompt, seed, resolution and steps.
Is FP16 Dense always faster on every Mac?
No. It is an optional computation path. Performance and memory use depend on the model and device, so compare both modes for your workload.