Krea-2-Turbo-Q4
Conversion status: all components complete.
Derived from krea/Krea-2-Turbo.
Source revision: 98e0fe118d17c9e3547fbb2e25acdbae2cadf7c7. Quantizer: MLX 0.32.3.
Linear and embedding matrices use affine 4-bit quantization with group size 64. Packed weights are U32; scales and offsets are BF16. Other tensors and the VAE retain their source precision.
| Component | Weight size (GiB) | Quantized matrices |
|---|---|---|
| text_encoder | 2.328 | 358 |
| transformer | 6.721 | 263 |
| vae | 0.473 | 0 |
Native Swift/MLX inference is available in zimageapp/zimage: KreaCLI --model /path/to/package --prompt "a red apple" --output out.png. Transformer, text encoding, LoRA and VAE decoding all run in Swift/MLX. The Python scripts remain independent conversion and numerical reference tools.
Krea2MLXPipeline is a descriptive model-index label, not a registered Diffusers pipeline.
Recommended sampling: 8 steps, CFG 0, fixed shift mu 1.15, resolution 1024–2048.
See conversion.json for source/output hashes and tensor counts; verification.json is written after the actual artifact audit.
Benchmark images and measured timings are separate artifacts; conversion alone does not establish image quality or speed.
The original Krea 2 Community License is preserved in LICENSE.pdf.
Original model documentation and attribution are preserved in SOURCE_README.md.
All recipients are bound by the original Krea 2 Community License in LICENSE.pdf. This is an unofficial quantized derivative. See Notice.
Quantized