Apple Silicon's native image generator, rebuilt line by line

What it does

MFLUX is a from-scratch rewrite of modern generative image models—FLUX, Z-Image, Qwen Image, FIBO, SeedVR2—using Apple’s MLX framework instead of PyTorch. It runs locally on Apple Silicon, supports quantization, LoRA fine-tuning, ControlNet, inpainting, upscaling, and depth estimation. The project explicitly mimics Karpathy’s educational style: minimal, explicit, line-by-line.

The interesting bit

The README openly admits the easiest way to navigate its sprawling CLI in 2026 is to ask Cursor or Claude Code. That’s either refreshing honesty or a quiet admission that seven model families with divergent flags have outgrown manual documentation. Either way, the ecosystem has noticed—there are already four third-party UIs built on top.

Key highlights

• Native MLX implementations, not PyTorch wrappers; only Hugging Face tokenizers are borrowed
• Supports distilled and base variants across models from 3B to 20B parameters
• Quantization, local model loading, multi-LoRA, and metadata export work across model families
• Model-specific tricks: FIBO’s JSON prompt understanding, SeedVR2 upscaling, Qwen’s world knowledge, Depth Pro from Apple
• Third-party frontends already exist for ComfyUI, Streamlit, FastHTML, and native macOS

Caveats

• No training support for FIBO, SeedVR2, Qwen Image, or Depth Pro (training only available for Z-Image and FLUX.2)
• FLUX.1 is explicitly marked “legacy” despite still being listed
• README contains a typo (“qaility”) that has survived at least one revision

Verdict

Mac-owning ML engineers who want local image generation without CUDA dependencies should look here. PyTorch die-hards on Linux or NVIDIA hardware gain nothing—there’s a DGX install path mentioned, but the entire value proposition is MLX-native Apple Silicon performance.