How to run Z-Image Turbo on a Mac.

Quick answer

Use Kynd’s 6.1 GB Z-Image Turbo 4-bit MLX pack on any supported Apple Silicon Mac with 16 GB or more unified memory. The streamed engine now uses that same checkpoint and the same 1024px square canvas at every supported tier. BF16 remains an explicit specialist route rather than an automatic hardware-dependent switch.

New: Read why Kynd rebuilt the heavy-memory path →

Z-Image Turbo is unusually approachable for a capable local image model. The official 6B Turbo release is distilled to eight model forwards; Kynd’s local request recipe exposes this as a nine-step generation.

The model and working tensors still share memory with macOS, the Kynd interface and any loaded chat model. Kynd uses a memory lease around generation and unloads the image service after idle time, so a local image session does not have to pin that memory forever.

Support and calibrated tiers reflect the Kynd build shipping on 14 August 2026.

Choose your unified memory

What should I run?

Which Z-Image model should you run?

For most Mac users, there is one simple answer: the 4-bit MLX pack. It is the route Kynd can provision, resume and support end to end.

4-bit is not merely the small-Mac option. In one Kynd M1 Max measurement at 512×512 and nine steps, the 4-bit MLX pack completed in about 55 seconds versus about 62 seconds for BF16. That is a single machine and workload—not a universal benchmark—but it is why Kynd recommends 4-bit even when both fit.

What Kynd currently supports.

Native MLX generation

Shipping — The preferred service runs on Apple Silicon and keeps the image path local. Staged phases, lazy transformer blocks and tiled decode hold the measured 1024px route to an 8.58 GB MLX peak. Q4 on every automatic tier.

In-app model setup

Shipping — The 4-bit model has visible progress, cancellation and resumable downloading. Kynd checks free disk space and keeps extra staging headroom. ≈ 6.1 GB + 3 GB headroom.

Prompt controls

Shipping — Use a seed, 1–12 steps and square or landscape sizes from 256 through 1344 pixels, in multiples of 64. 1024 × 1024 · 9-step default.

LeMiCa cache

Optional — The native route exposes slow, medium and fast cache modes for users who want to trade a little iteration behaviour for speed. Native MLX path.

BF16 MLX

Detected — Kynd can select a complete compatible BF16 model that you converted locally, but does not currently offer a one-click BF16 download. 32 GB minimum.

Image editing

Not in this path — This Kynd service is currently prompt-to-image. Do not expect the Z-Image route to edit or inpaint an uploaded frame. Text to image today.

The practical settings.

Start here: 1024 × 1024 · 9 steps

• The official Turbo recipe.
• Guidance stays at zero.
• Use a fixed seed when comparing prompts.
• Kynd unloads other local models before the full 1024px route begins.

Supported envelope: 256–1344 pixels

• Width and height use multiples of 64.
• Steps can range from 1 through 12.
• Very wide or tall canvases cost more memory.
• Free chat-model memory before large renders.

How to install and run Z-Image in Kynd.

1. Check the Mac. Open Apple menu → About This Mac. The supported Kynd route requires Apple Silicon, macOS 14 or later and at least 16 GB unified memory.
2. Open the local image setup. Start an image generation or open Kynd’s model settings. If Z-Image is missing, Kynd presents the local model card and the recommended 4-bit route.
3. Install the native runtime and model. Choose Install. Kynd downloads its pinned MLX runtime and the 6.1 GB 4-bit weights with progress you can cancel or resume.
4. Begin with the default recipe. Use 1024×1024, nine steps and guidance zero. Kynd stages the model phases and unloads other local models automatically.
5. Keep the seed when iterating. A fixed seed makes prompt comparisons meaningful. Change one phrase at a time, then raise the canvas size after the composition works.
6. Let Kynd return the memory. The local service unloads after its idle window so Metal memory becomes available to chat, code and video again.

Z-Image on Mac FAQ.

Can a 16 GB M1 or M2 Mac run Z-Image Turbo?

Yes. Kynd’s streamed 4-bit MLX route has a 16 GB unified-memory floor and keeps the normal 1024px square canvas. Its constrained-memory engineering run completed at a measured 8.58 GB MLX peak; wider physical 16 GB mileage remains part of release validation.

Should I use 4-bit or BF16?

Use 4-bit unless you have a specific evaluation reason to compare BF16. It is the supported in-app installation, uses much less disk and memory, and was slightly faster in Kynd’s current M1 Max measurement.

Can Kynd install the BF16 model for me?

Not currently. The BF16 route is detected-only because Kynd does not have a published compatible pack to install automatically. Advanced users can convert the upstream weights to the expected MLX layout.

Does everything stay on my Mac?

Generation runs against Kynd’s localhost image service and local weights. The service is authenticated locally, and Kynd unloads it after idle time to release unified memory.

Where can I verify the upstream model?

See the official Z-Image Turbo model card for architecture, capabilities and terms, and the native MLX implementation used by Kynd’s pinned local path.

Also useful: Run LTX 2.3 or 2.5 locally on Mac → · Choose a Qwen 3.8 27B quant for Mac →