⚡ Krea2 Turbo BF16 → W6A8 — NEW!

Experimental 6-Bit Quantization | Created with VINCE W6A8 Studio

A new quantized version of the original Krea 2 Turbo BF16, converted by Vince AI using the standalone VINCE W6A8 Studio.

This version introduces W6A8 quantization (6-bit weights / 8-bit activations), offering a different balance between model size and numerical precision compared to W4A4 and W4A8.

Key Features

• W6A8 quantization with ConvRot.

• 240 layers successfully quantized.

• Direct conversion from the original BF16 checkpoint.

• Clean base model — no merged LoRAs.

• Successfully tested with Forge Neo.

• Convincing visual results at 6 and 8 steps.

• Supports additional LoRA experimentation.

W4A4 vs W4A8 vs W6A8

• W4A4 — Low VRAM Edition: Maximum focus on memory efficiency.

• W4A8 — Balanced Edition: 4-bit weights with higher activation precision.

• W6A8 — Enhanced Precision Edition: 6-bit weights and 8-bit activations, offering greater weight precision.

Recommended Settings

• 6 Steps — Recommended

• 4 Steps — Fast Mode

• 8 Steps — Alternative for more complex scenes

• VAE: Wan 2.1

• Text Encoder: Qwen3VL 4B FP8

Credits & License

Original model developed by Krea.ai.

W6A8 conversion by Vince AI, using VINCE W6A8 Studio.

Unofficial community conversion, not affiliated with or endorsed by Krea.ai.

Original model: https://huggingface.co/krea/Krea-2-Turbo

Official licensing information: https://www.krea.ai/krea-2-licensing

Distributed subject to the applicable Krea 2 Community License terms.

Experimental W6A8 format. Actual VRAM usage, speed and image quality may vary depending on hardware and software configuration.