Same speed if you have enough vram but yes reduces vram like %20-30, so that's why its much faster if you run out of vram.  Quality wise i couldn't notice much difference it is still great.

Error occurred when executing CheckpointLoaderNF4:

load_checkpoint_guess_config() got an unexpected keyword argument 'model_options'

File "G:\ComfyUI\execution.py", line 152, in recursive_execute
output_data, output_ui = get_output_data(obj, input_data_all)
File "G:\ComfyUI\execution.py", line 82, in get_output_data
return_values = map_node_over_list(obj, input_data_all, obj.FUNCTION, allow_interrupt=True)
File "G:\ComfyUI\execution.py", line 75, in map_node_over_list
results.append(getattr(obj, func)(**slice_dict(input_data_all, i)))
File "G:\ComfyUI\custom_nodes\ComfyUI_bitsandbytes_NF4\__init__.py", line 178, in load_checkpoint
out = comfy.sd.load_checkpoint_guess_config(ckpt_path, output_vae=True, output_clip=True, embedding_directory=folder_paths.get_folder_paths("embeddings"), model_options={"custom_operations": OPS})
关闭

I confirm what the guy wrote above uses less ram. But when you have enough ram for fp8 the difference for me is 0-2 seconds. On the plus side it generates very similar images, differences in detail.

worked just fine last night, today:

Error occurred when executing CheckpointLoaderNF4: Expecting value: line 1 column 1 (char 0) File "D:\COMFY\ComfyUI\ComfyUI_windows_portable\ComfyUI\execution.py", line 152, in recursive_execute output_data, output_ui = get_output_data(obj, input_data_all)

other error messgae here but the above is problaby the key

wolf