MisterSnowCrash/nitrofusion2-mlx-q4

NitroFusion2, MLX 4-bit (group size 64)

A 4-bit MLX quantization of ChenDY/NitroFusion2, a 1-2 step distillation of Z-Image-Turbo, for on-device inference on Apple silicon (Mac, iPhone, iPad) with MLX Swift. 1 or 2 steps, FlowMatchEuler shift 6.0, no classifier-free guidance.

What was changed (Apache 2.0, section 4b)

• Transformer: the NitroFusion2 transformer weights (Apache 2.0) with every Linear layer quantized to 4-bit affine, group size 64 (the same modules and dtypes as the Z-Image-Turbo q4 tier below), split into two shards. Norms and pad tokens are unchanged.
• Text encoder (Qwen3-4B), tokenizer and VAE: copied unmodified from the q4/ tier of SceneWorks/z-image-turbo-mlx at revision bb2bc9893b3c49ae96c813350775f791a2e8bc80 (Apache 2.0, "unmodified-weights re-host, re-quantized for MLX"). NitroFusion2 uses the same encoder, tokenizer and VAE as Z-Image-Turbo.
• scheduler/scheduler_config.json: shift set to 6.0 (NitroFusion2's schedule).
• No fine-tuning, no change to the training data or to the model's behavior beyond quantization error.
• Size: about 5.9 GB (5.5 GiB).

License and notices

• Licensed under the Apache License 2.0, the same license as the original models (see LICENSE).
• NitroFusion2: Dar-Yen Chen, Yi-Zhe Song (University of Surrey), arXiv:2611.05648, https://huggingface.co/ChenDY/NitroFusion2 (Apache 2.0).
• Z-Image-Turbo: Tongyi-MAI (Alibaba), https://huggingface.co/Tongyi-MAI/Z-Image-Turbo , code and license: https://github.com/Tongyi-MAI/Z-Image (Apache 2.0). Paper: Z-Image Team, arXiv:2511.22699.
• Text encoder: Qwen3-4B by Alibaba Cloud (Apache 2.0), https://huggingface.co/Qwen/Qwen3-4B
• VAE: the FLUX.1 autoencoder by Black Forest Labs; "The weights of the autoencoder are also released under apache-2.0" (https://github.com/black-forest-labs/flux).

Use of the model must follow applicable law; the original authors' model cards apply.