Core AI is Apple's on-device ML runtime in iOS 27 / macOS 27 and the successor to Core ML: PyTorch models are exported with Apple's coreai-torch (LLMs: coreai.llm.export) into .aimodel bundles that run on the GPU or the Neural Engine, e.g. Qwen3-8B 4-bit decodes at 94 tok/s on an M4 Max GPU, MLX 90 under the same protocol (apple-silicon-llm-bench, macOS 27 beta, 2026-06).

Z-Image-Turbo — Core AI (macOS)

Alibaba Tongyi-MAI Z-Image-Turbo (6B, Apache-2.0) — a Single-Stream Diffusion Transformer (S3-DiT) — converted to Core AI and generating images entirely on the Mac GPU.

A Qwen3-4B text encoder conditions a 34-block DiT that denoises in 8 FlowMatchEuler steps with classifier-free guidance; a 16-channel VAE decodes. Photoreal by default.

One DiT graph covers 256², 512² and 1024², and any prompt length — both the image-token and caption axes are dynamic, at a ~5–9 % cost over static shapes.

What's here

Weights are bf16; the graph boundaries are fp32 (--io-fp32). That is not a preference: a Swift host cannot fill or read a bfloat16 NDArray, and bf16 is the only dtype this DiT is numerically safe in. Casting at the boundary costs ~15 % per forward and improves fidelity (PSNR 39.5 → 42.6 dB) because nothing rounds on the way in.

The glue/ is what keeps a host from re-implementing the reference: RopeEmbedder is a per-axis table lookup, so three tables reproduce it exactly at any resolution and prompt length, and the timestep MLP ships as a 2 MB graph.

bf16, not int8. On this compute-bound graph weight-only int8 is slower than bf16 (2.35 vs 0.89 s/forward at 512²) because it dequantizes back to 16-bit and runs the same matmul — it only wins on bandwidth-bound shapes and on footprint. bf16 is also what keeps the port numerically near the fp32 reference.

Speed & fidelity (M4 Max, vs the fp32 diffusers reference)

Per-step velocity correlation vs the reference is ≥ 0.9997 at every step and both CFG branches. PSNR is not comparable across prompts: a texture-heavy oil-painting prompt scores 27.7 dB while being visually indistinguishable from the reference.

Usage

Run it in CoreAIImageGen (macOS): pick "Z-Image-Turbo 512" or "… 1024" → Download & Load → Generate. The host loop is ZImagePipeline.swift; the Python twin is conversion/zimage/pipeline_engine.py. Both agree with the fp32 reference to ~42.6 dB.

The DiT graph takes host-prepped inputs (patchify, RoPE, pad masks) and returns the velocity; the sampler loop lives on the host:

# per step, for cond and uncond:
#   ins = build_native_inputs(rm, latent, cap)          # patchify + RoPE + pad masks
#   v   = dit(**ins, adaln=t_embedder(t * t_scale))     # Core AI graph
#   vel = unpatchify(v[:, :n_img])
# noise_pred = -(pos + guidance * (pos - neg))          # Z-Image CFG is NEGATED
# latent    += dsigma[s] * noise_pred                   # FlowMatchEuler
# image      = vae(latent)                              # unscale: z/0.3611 + 0.1159

Three details each cost a wrong image:

1. the DiT conditions on the encoder's penultimate hidden state (hidden_states[-2]);
2. the CFG is negated: -(pos + g·(pos − neg)), not neg + g·(pos − neg);
3. captions are padded to a multiple of 32 with a learned pad token that is real attention context — n_cap = round_up(L, 32) must match, and cond/uncond generally have different n_cap (hence the dynamic caption axis).

Notes

• macOS only. fp16 sends this DiT all-NaN at sampler step 2 (depth-driven, at every resolution); bf16 is exact — and coreai-build compile refuses a bf16 module, which iOS needs for graphs this size. Full analysis in the port notes.
• The text-encoder graph is fixed at 64 chat-templated tokens (≈ 35–40 words).
• guidance=0 skips CFG — half the work, a different composition, still clean at 256².
• Weights are not redistributed as source: the graphs are produced from the original Apache-2.0 checkpoint by the conversion scripts in the zoo.

License: Apache-2.0 (inherited from Z-Image-Turbo).

---

More models in this format: Core AI Model Zoo — 75 models, each with the recipe that produced it.

Want a different model on-device? Open a request — free, open weights only; the export and its measured numbers get published publicly.