← sources

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

source: reddit (reddit:LocalLLaMA)published 2026-09-20 18:10ingested 2026-09-21 02:38

retained text

After seeing a few posts on here about it, I finally tried exl3 3bpw and exllamav3 for running flash next - with amazing results. 

  
On 3x3090s, 128GB DDR4: 1500 prefill, 80 tps decode 

On 1x5090, 128G. DDR4: 1500 prefill, 29 tps decode

  
Both at 262k context, both with vision/spec decoding. Really impressed, definitely replacing vllm/llama.cpp for me on this model.

  
Quant capacity seems good so far, going to gest the 4bpw later for comparison.

  
Check it out if you were sleeping on it like I was!

document fetched 2026-09-21 00:36 · canonical https://reddit.com/r/LocalLLaMA/comments/1wlo9nz/one_more_you_should_try_exllamav3exl3_for_flash