Running Z-Image-Turbo natively on Mac (Swift/Metal, no Python)
#166
by gwenn-ha-dev - opened
Hi! I've built a free, open-source macOS app that runs Z-Image-Turbo natively on Apple silicon, in Swift and Metal: no Python, no ComfyUI, nothing to wire. Pick the model in the app, it downloads and converts the weights, then you type a prompt. Text-to-image, image-to-image and LoRAs (Civitai, Comfy, kohya and diffusers layouts) all work, and fine-tuned DiTs import by drag and drop.
- Checked against the reference: every stage (Qwen3 text encoder, transformer, VAE) is compared channel by channel with the diffusers implementation in fp32; 1.2e-5 at 1024².
- On an M1 Pro with 16 GB: ~82 s at 1024², 8 steps, with a 3.8 GB peak and zero swap at 1024×1536.
- Smaller Macs: besides the full weights (20.4 GB), it can install 8-bit (11.7 GB, unsloth GGUF and Disty0's SDNQ int8) or GGUF Q4_K_M (9.5 GB, unsloth) versions, kept at their published precision. The 4-bit version is preselected on 8 GB Macs.
Apache-2.0, notarized DMG: https://github.com/gwenn-ha-dev/Siliconed
happy to answer questions. Thanks to the Tongyi-MAI team for releasing the Turbo under Apache 2.0!