wabibito/Krea2-Onyx-Turbo-mlx-4bit

Krea 2 Onyx Turbo mlx 4bit

A 4-bit MLX quantization of the official krea/Krea-2-Turbo, built for on-device inference in the Onyx app with a documented, reproducible pipeline.

• Source: official krea/Krea-2-Turbo (diffusers fp16), quantized in-house.
• Quantization: 4-bit, group size 64, MLX. Only Linear leaves are quantized; norms / embeddings / convs / the VAE stay full precision. Components: transformer/ (Krea2Transformer2DModel, 28 blocks + text-fusion), text_encoder/ (Qwen3-VL), vae/ (AutoencoderKLQwenImage), tokenizer/, scheduler/.
• Inference: distilled — ~8 steps, CFG off.
• Verified: reloaded through the Onyx engine and render-checked (coherent image, not noise) before publish.

This is a quantized derivative of Krea 2 for efficient on-device use; all credit for the model to Krea AI.

License

Krea 2 Onyx builds are Derivative Models of Krea 2 and are distributed under the Krea 2 Community License Agreement (LICENSE.pdf in this repository; https://krea.ai/krea-2-licensing). By downloading or using these weights you agree to be bound by that Agreement, including:

• its prohibited uses and Acceptable Use Policy;
• the requirement that deployments implement content filters (prompt/output filtering, abuse detection, moderation) to prevent prohibited uses and outputs;
• its Commercial Use terms (permitted under the Agreement only below the revenue threshold in s2.3; above it, an Enterprise License from Krea is required);
• passing these same obligations to anyone you redistribute to (s3.1).

This is an unofficial community build - not an official Krea product, and not endorsed by or affiliated with Krea.ai, Inc. See NOTICE for attribution.