This is a mixed precision INT8 Tensorwise checkpoint. It is ready for use in ComfyUI with the following custom node:

https://github.com/BobJohnson24/ComfyUI-Flux2-INT8

It achieves 2x performance on a 3090 over fp8 and bf16. On newer cards it will not be as much of a performance increase.

On a 5060 I see a 1.3x speed-up compared to FP8.