A Zig/Metal inference engine for diffusion models on Apple Silicon: one small binary with its own Metal kernels, no Python, PyTorch or model server. It runs two open-weight text-to-image models locally, FLUX.2 Klein 4B (distilled and base) and Z-Image-Turbo, for generation and editing.

1024×1024, four steps, M4 Max / 128 GiB, cached weights. Playback at 1×. Full reroll/save demo.

zDraw exists for two properties most engines do not offer. Renders are deterministic: the same request produces the same bytes, so an image can be reproduced from its recipe, a benchmark card from another Mac is comparable with yours, and a kernel change is judged by a hash rather than by eye. Memory is treated as a budget rather than a side effect: weights stay memory-mapped and activations live in reused pools, so a 1024×1024 Klein render adds about 3 GiB of process memory on top of the weight file and the engine can share a Mac with other work.

• Generate from text, edit a photo by instruction, redraw a masked region, vary an image with --init-image --strength, or render several seeds at once.
• Klein 4B in four steps, its undistilled base checkpoint in 50 guided steps, and Z-Image-Turbo with a product or bit-exact strict decoder profile.
• Watch Klein's approximate previews as it samples, then see the decoded image.
• Keep models loaded in a session; reroll, undo, switch models and save results. Stanza provides terminal history and completion.
• PNGs from generate and bench carry a recipe chunk: model, prompt hash, seed, size, steps and guidance.
• Resumable downloads, W16 packing with optional W6/W4 quantisation and LoRA merging, and a prompt and image safety filter. zdraw doctor reports the routes your Mac will use; zdraw bench renders the certified case.

What is in the repository

• src/: the engine, one folder per owner: metal/ (kernels, contexts and the Objective-C bridge), klein/ and zimage/ (the two model families), sampling/, vae/, text/ (the Qwen3 encoder and tokenizer), pack/ (weight files and sidecars), runtime/ (model state, profiles, metrics), cli/ (commands, sessions, terminal previews, safety) and control/ (experimental execution controls). main.zig is the CLI root, lib.zig the package the tools import, tests.zig the test root.
• cmd/: one file per tool executable: the pack builders (kleinpack, zpackbuild) and the kernel benchmarks and gates; zig build check compiles them all.
• certified/hashes.json: the reference hashes that zdraw bench checks.
• docs/: usage, features, architecture, the measurement protocol, the validation record and the resolution report with its evidence.
• tools/, vendor/, safety/, community/: the quality and benchmark harnesses, the vendored Metal attention kernels (MLX steel and MFA, MIT), the prompt rules, and the table of benchmark cards from other Macs.

Model weights are not included; zdraw fetch downloads and packs them (Klein about 24 GB on disk, Z-Image about 49 GB). There is no desktop app or SDK in this repository.

Quick start

You need an Apple Silicon Mac with Zig 0.16.0 and Xcode's Metal compiler. Everything below was validated on an M4 Max with 128 GiB running macOS 26; other Apple Silicon Macs build and pass the weight-free tests in CI, but their generation speed, memory and output are unvalidated.

xcrun metal --version
zig build
./zig-out/bin/zdraw fetch flux2-klein-4b
./zig-out/bin/zdraw generate --model flux2-klein-4b \
  --prompt "a red fox in deep snow" --show --out fox.png

Usage and troubleshooting · Supported features · How it works, a short tour of the model and the implementation

Speed and memory

1024×1024, four steps, W16 weights, M4 Max / 128 GiB, measured 5 October 2026:

Model	End-to-end time	Peak RSS	Peak footprint
FLUX.2 Klein 4B	11.96 s	2.33 GiB	3.02 GiB
Z-Image-Turbo	24.52 s	3.95 GiB	4.72 GiB

Median of three fresh processes with cached weights, including loading through PNG saving. Memory is the maximum measured across those runs, not minimum Mac RAM: mapped weight pages can occupy additional memory. All tested resolutions, image samples and methods · Comparisons with mflux, iris.c and diffusers on the same Mac.

Editing

“Add a blue wool scarf around the fox's neck”:

Before	After

Run this edit. Instruction edits can change other areas; masks give spatial control with a feathered boundary.

Contributing

Bug fixes, documentation and results from other Macs are welcome: run zdraw bench and submit the card to the community benchmark table. See CONTRIBUTING.md for setup and tests.

Licence

Apache-2.0, including commercial use. Third-party code is listed in NOTICE; model weights retain their upstream licences.