An MLX q8 quantization of krea/Krea-2-Raw, the Krea 2 base checkpoint, for training LoRAs with mflux without the RAM to quantize the full-precision model locally. Krea recommends training on Raw; Turbo is the distilled inference checkpoint.

Quantized with mflux-save (q8). The text encoder is kept at bf16 (quantizing it degrades conditioning).

Point an mflux-train config at this repo ("model": "krea-2-raw", "model_path": "<this repo>"), then use the trained LoRA on Krea 2 Turbo at inference.