Onyx-Qwen-Image-Edit-2509-mlx

Onyx · Qwen-Image-Edit-2509 (MLX, 8-bit DiT)

Onyx's own MLX export of Qwen/Qwen-Image-Edit-2509 (Apache-2.0) — the image EDIT model: a source image plus an instruction ("make the sky purple", "remove the sign") in, the edited image out. Built for the Onyx on-device engine (OnyxDiffusion), which streams the DiT block-by-block so the denoise memory footprint stays flat on iPhone.

• transformer/ — QwenImageTransformer2DModel (60 blocks), uniform 8-bit group-64. 4-bit was render-verified to add uniform fine grain (per-step error accumulates over 60 blocks x 50 steps); 8-bit is indistinguishable from full precision.
• text_encoder/ — Qwen2.5-VL-7B conditioning tower, mixed precision (default 4-bit group-64 with the quant-sensitive attention Linears at 8-bit; each Linear's bits are self-described by its own weight/scales/biases shapes).
• vae/ — AutoencoderKLQwenImage, unquantized bf16.
• tokenizer/, processor/, scheduler/, model_index.json — unchanged from the official source.
• sha256.json — per-file integrity manifest.

~27 GB total. License: Apache-2.0 (inherited from the base model).