flux2-dev-turbo-Q4_K_M

FLUX.2-dev + Turbo (Merged) - GGUF[Q4_K_M]

Overview

This repository contains a merged and quantized version of the 32B FLUX.2-dev model with the Turbo LoRA baked directly into the UNET weights. It is quantized to Q4_K_M format using a custom llama.cpp build specifically patched for the FLUX architecture.

Why use this merged version?

If you use a base GGUF model and apply a LoRA node dynamically in ComfyUI, the engine applies "lowvram patches" (dequantizing layers on the fly during generation). For a 32B model, this drastically kills generation speed, even if the model fits in your VRAM.

By using this pre-merged GGUF file:

• 0 LowVRAM patches applied.
• The model fits entirely into 24GB VRAM (full load: True on RTX 3090/4090).
• Maximizes performance when paired with SageAttention or FlashAttention.

File Details

• Base Model: FLUX.2-dev (32B parameters)
• LoRA Applied: FLUX.2-dev-Turbo (Weight: 1.0)
• Format: GGUF
• Quantization: Q4_K_M
• File Size: ~17.8 GB

How to use in ComfyUI

1. Download the flux2-dev-turbo-Q4_K_M.gguf file.
2. Place it in your ComfyUI/models/unet/ directory.
3. Use the Unet Loader GGUF node to load the model.
4. ⚠️ IMPORTANT: DO NOT use a Load LoRA node for the Turbo LoRA. The weights are already baked into this UNET. Just connect it straight to your Sampler.
5. Setup your Sampler for Turbo (e.g., 8 steps, CFG 1.0-1.5, depending on the Turbo LoRA requirements).

Hardware Requirements

• RAM: 32 GB+ recommended
• VRAM: ~24 GB (Fits comfortably on RTX 3090 / 4090 alongside a quantized Mistral-3 encoder).

Credits & Acknowledgements

• Base model by Black Forest Labs.
• Turbo LoRA by Fal.
• Quantized using the awesome ComfyUI-ModelQuantizer and city96's patched llama.cpp.

Note: Please adhere to the original FLUX.2-dev non-commercial license when using this model.