One more 'you should try ExllamaV3/exl3 for flash next' appreciation post

A user shared their positive experience with ExllamaV3/exl3 3bpw for running "flash next," noting impressive results. On a setup with 3x3090s and 128GB DDR4, they achieved 1500 prefill and 80 tps decode. A single 5090 with 128G DDR4 yielded 1500 prefill and 29 tps decode. Both configurations operated at 262k context with vision/spec decoding, leading the user to replace vllm/llama.cpp with ExllamaV3 for this model. They also plan to test the 4bpw version.