# Best VLLM setup for Qwen3.8 Flash Next on 2x5090 + 128GB of RAM? — retained record

> Retained reddit record from reddit:LocalLLaMA.

Canonical: https://highsignal.app/data/records/3bbcfb3c63d2fffe

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:LocalLLaMA)  published 2026-10-21 05:57 ingested 2026-10-22 02:43

[original source ↗](https://www.reddit.com/r/LocalLLaMA/comments/1xbf4vp/best_vllm_setup_for_qwen38_flash_next_on_2x5090/)

## retained text



```
Hi all,

I am running Qwen3.8-27B for a local setup with 2x5090s where the use case is mostly agentic programming. For this, VLLM is the only framework that does not trash the cache and allows for efficient use. I am currently using Qwen3.8-27B-INT6-Mixed-AutoRound from Minachist with pretty good results. 

But I am considering to try Flash Next on this rig, and while I've seen lots of discussions with different framewors, and many new vibe-coded ones, I haven't found a good balanced recipe for 2x5090s, flash-next and vllm. It seems pretty different to serve than Qwen3.8-27B where you fit everything on VRAM, so I'd like to know what are people doing. If you have experience with this kind of setup, could you share it?

Thanks!
```



document fetched 2026-10-22 00:45 · canonical https://reddit.com/r/LocalLLaMA/comments/1xbf4vp/best_vllm_setup_for_qwen38_flash_next_on_2x5090
