Onyx Z Image — 3-bit MLX

• Source: official Tongyi-MAI/Z-Image (diffusers), quantized in-house.
• Quantization: 3-bit, group 64, MLX. Only transformer + text-encoder Linear leaves are quantized; norms/embeddings/convs and the whole VAE stay fp16. Components: transformer/, text_encoder/ (Qwen3), vae/ (16-ch FLUX-family), tokenizer/, scheduler/.
• Inference: ~22-28 steps + real guidance (undistilled). Built as MIXED 3/4-bit: quant-sensitive layers (embedders, final layer, adaLN modulation, attention) stay 4-bit; the FFN bulk is 3-bit. Uniform 3-bit mottles the undistilled DiT; this recovers a clean image at near-3-bit size.
• Verified: reloaded through the Onyx engine and render-checked before publish.