Mage-Flow-Edit-Turbo-MLX-Serve-8bit

Mage-Flow-Edit-Turbo-MLX-Serve-8bit

8-bit mirror of microsoft/Mage-Flow-Edit-Turbo for mlx-serve. Half the download, half the memory, same distilled 4-step schedule.

Weights: 9.7 GB (upstream bf16: 17.5 GB).

Judged against bf16 at the same seed on text-to-image, single-reference edits and multi-reference composition: no visible quality difference.

Run it

Download MLX Core.app, open the Image tab, and pick Mage-Flow Edit Turbo 8-bit from the model menu. It downloads with a progress bar and generates in the same window. No terminal, nothing to configure.

Prefer Homebrew? It is a third-party tap, so tap it first:

brew tap ddalcu/mlx-serve https://github.com/ddalcu/mlx-serve
brew install --cask mlx-core

Drop in a source image, type an instruction, generate. Add more references and it composes across them.

Driving it from code instead? The app runs the server on http://localhost:11234, so POST to /v1/images/edits. That is the OpenAI image-edit shape, so the official SDK's client.images.edit() works as-is, with repeated image[] for multi-reference.

mlx-serve

A native Zig server for Apple Silicon. No Python, no cloud, no Electron. One 9 MB binary.

• One server, every modality. Chat, images, video, music, speech with voice cloning, and 3D, all running natively on MLX.
• Points at what you already use. OpenAI- and Anthropic-compatible APIs on http://localhost:11234, so Claude Code, the OpenAI SDK, Continue, Cursor and Open WebUI just work.
• Any LLM, not just these. Every MLX model and every GGUF on Hugging Face, with speculative decoding built in.
• MLX Core.app included. Signed macOS menu-bar app: chat, agent mode with MCP tools, model downloads, and every generator above, no terminal needed.

mlxserve.com · GitHub

If it is useful to you, a star on GitHub genuinely helps.

Recipe

DiT and text-encoder linears affine-quantized at 8-bit, group size 64, via mlx.core.quantize. Left at bf16: the whole VAE (its precision is load-bearing for a distilled 4-step model), the token and position embedding tables (they are gathered, not matmul'd), and any linear with a dimension under 512.

Dropped because the engine never loads them: the VAE's training-time encoder half.

Dropped too, the repo files it never opens: chat_template, tokenizer_config, generation_config, video_preprocessor_config, vocab.json + merges.txt, assets.

Built by tests/convert_mageflow_weights.py in the mlx-serve repo. Original model and weights by Microsoft, MIT licensed.