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).

GLM-Image — Core AI

ZhipuAI's GLM-Image (16B, MIT) converted to Core AI for on-device image generation on Apple Silicon (macOS 27+) — the zoo's first autoregressive + diffusion hybrid.

GLM-Image generates in two stages: a 9B GLM-4 autoregressive model writes the image as a grid of discrete visual prior tokens (sampled left-to-right like an LLM, ~36 tok/s on M4 Max), and a 7B flow-matching DiT denoises the actual pixels conditioned on those tokens, followed by a 16-channel VAE decode. Both native 1024×1024 and a faster 512×512 variant are included.

macOS only. The AR (9.6 GB) + DiT (9.2 GB) weights are resolution-independent, so the 512 variant is no smaller — an iPhone port needs int4 + sequential component loading.

Components

Weights are int8 (per-block-32) for AR/DiT; the VAE stays fp32 (fp16 overflows its activations). Compute precision float16 on GPU. One resolution needs ~19.7 GB on disk (AR + one DiT + one VAE).

Usage

Sample app (easiest)

CoreAIImageGen (macOS) — run the CoreAIImageGenMac scheme, pick GLM-Image 1024 (AR+diffusion) from the model menu, tap Download & Load, type a prompt, Generate.

• Steps 20 (default) ≈ 4.5 min/image on M4 Max — reference quality
• Steps 12 ≈ 2.5 min — nearly indistinguishable
• The 512 variant generates in ~1 min

The AR stage is sampled (temperature 0.9 / top-p 0.75, the upstream generation_config), so re-rolling the seed meaningfully changes composition.

Pipeline contract (build your own)

The hybrid does not fit Apple's high-level CoreAIDiffusionPipeline; the host loop is ~300 lines of Swift (see the app's GlmImagePipeline):

1. AR: tokenize(prompt) + fixed grid suffix → S=1 prefill/decode with host-computed 3D-mRoPE cos/sin → sample visual tokens → 2× upsample the large grid → prior[1, N] (N = 1024 @512, 4096 @1024).
2. DiT: 20-step CFG (guidance 1.5, prior_scale 1/0 for cond/uncond). Condition the DiT on the raw integer schedule trunc(linspace(1000,1,steps+1)) − 1; step the latent with the mu-shifted sigmas σ' = μ/(μ + (1/σ − 1)), μ = 0.75·side/256 + 0.25.
3. VAE (fp32, CPU): latent → RGB.

⚠️ Feeding the DiT the shifted timesteps (sigma×1000) instead of the raw schedule skews the adaLN time conditioning every step and drifts colors — resolution-dependently (mild at 512, strong at 1024). This is the one non-obvious contract in the port.

Limitations

• Glyph rendering (text in images) is not wired: prompts containing quoted text ('…', "…", 「…」) route through a T5 glyph encoder that is not part of this bundle. Keep prompts glyph-free.
• Negative prompts are a no-op (GLM's CFG drops the prior tokens instead).
• Image-to-image / editing (GlmImageKVCache) not yet exported.

Quality

Verified against the bf16 diffusers reference: DiT single-forward cosine 0.9999, VAE byte-exact, AR prior-injection reproduces the reference composition; end-to-end output is visually on par with GlmImagePipeline at the same step count.

License

MIT (inherited from zai-org/GLM-Image).

---

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.