Krea-2-Turbo-MLX-Serve-mixed-4-8

Krea 2 Turbo — MLX-Serve (mixed-4/8)

A single-download, ready-to-run bundle of krea/Krea-2-Turbo — a 12.9B single-stream MMDiT text-to-image model — packaged to run natively on Apple Silicon in mlx-serve. Everything the engine needs is in this one repo: the quantized transformer, the Qwen-Image VAE, the Qwen3-VL-4B text encoder, the tokenizer, and a ready config.json. No gated downloads to assemble, no Python.

Independent, unofficial port. Not an official Krea product and not endorsed by Krea. A modified derivative of krea/Krea-2-Turbo under the Krea 2 Community License.

Example output — photorealistic 1024² portrait, generated in-app via mlx-serve on Apple Silicon (8-step Turbo, no guidance).

🚀 Made for mlx-serve

mlx-serve is a native Zig server that runs MLX-format models on Apple Silicon and exposes OpenAI- and Anthropic-compatible HTTP APIs — chat, vision, embeddings, and native image / audio / video generation — all in one process, no Python. Image generation (FLUX.2 and Krea-2) runs through the same server as your LLMs, sharing one port and one memory budget.

# pick "Krea 2 Turbo" in the mlx-serve app's Image tab — it downloads this repo and generates.
# or headless, by path:
mlx-serve --serve --model-dir ~/.mlx-serve/models &
curl localhost:8080/v1/load-model -d '{"model":"<path-to-this-repo>"}'
curl localhost:8080/v1/images/generations \
  -d '{"model":"Krea-2-Turbo-MLX-Serve-mixed-4-8","prompt":"a red fox in the snow","size":"1024x1024","steps":8}'

→ mlx-serve on GitHub: https://github.com/ddalcu/mlx-serve — ⭐️ please star the repo to support the project!

What's inside

~15 GB total. Generation is 8-step flow-matching, no guidance (distilled Turbo). Sizes: any multiple of 16 in [256, 2048].

Quality

Validated numerically faithful to the pure-MLX reference (which is itself cos ≈ 1.0 vs the PyTorch original): full-pipeline pixel cosine 0.9996 at 512²/8-step. The 8-bit text encoder is near-lossless end-to-end (the conditioning washes through the diffusion transformer); the transformer is the near-lossless mixed-4/8 build (≈9.8 GB, vel-cos ≈0.998 vs bf16).

License & attribution

By downloading or using this repository you agree to be bound by the Krea 2 Community License Agreement (a verbatim copy is in LICENSE; attribution in NOTICE). In particular:

• Naming. Per the license, derivative model names begin with "Krea".
• Commercial use is permitted only if your total annual revenue is under $1,000,000 USD; otherwise you need an enterprise license from Krea ([email protected]).
• Content filtering. You must implement reasonable content-filtering safeguards in any deployment, and disclose AI-generated content where required by law. mlx-serve does not apply a safety filter by default — that obligation is yours.
• Acceptable Use Policy. Your use must comply with Krea's Acceptable Use Policy (incorporated by reference into the license).
• Not endorsed. Independent port; not an official Krea product and not endorsed by Krea.

The VAE, text encoder, and tokenizer are Qwen models used under Apache-2.0 (see NOTICE). You own the images you generate (subject to the license); Krea claims no ownership of outputs.

Credits

• Krea.ai — Krea 2 (base model & reference code)
• Qwen — Qwen-Image VAE & Qwen3-VL-4B text encoder (Apache-2.0)
• avlp12 — the validated pure-MLX Krea-2 port this bundle's transformer quantization derives from
• mlx-serve — the native Apple-Silicon server this bundle is packaged for. ⭐️ If this is useful, please star https://github.com/ddalcu/mlx-serve!