# Please recommend Qwen3.8-Flash-Next llama parameters — retained record

> Retained reddit record from reddit:LocalLLaMA.

Canonical: https://highsignal.app/data/records/b4f05f7de7d0fd9e

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


source: [reddit](https://highsignal.app/data/reddit) (reddit:LocalLLaMA)  published 2026-10-11 21:11 ingested 2026-10-12 02:39

[original source ↗](https://www.reddit.com/r/LocalLLaMA/comments/1x3jnaw/please_recommend_qwen38flashnext_llama_parameters/)

## retained text



```
I like tinkering with stuff and I think that now I understand at least on surface-level how a dense model like 27B works and what needs to be finetuned for a specific configuration.

Then, like everyone else, I discovered Strata and running a large moe locally became possible.

The problem is that I cannot get *anywhere close*  to Strata's speeds with vanilla llama server. I have no idea what most of the moe-specific flags actually do to my machine and it all feels like a huge black box without feedback. It bothers me to not understand what's happening inside and what I need to improve or upgrade to make it better. 

I'm trying to run an IQ3\_XXS on vanilla llama.cpp for which Strata gives about 35 tps on thinking and 50 tps code.

My best run so far on llama is 15 tps.

Hardware: 2 x RTX3060 12GB + RAM 64GB.

I cannot find the specific 800MB mtp q2\_0 drafter used by Strata on HF. The only one I found is the mtp-Qwen3.8-Flash-Next-Q4\_K\_M.gguf, which is 3.5x the size. *It makes decode considerably slower*. For now I'm testing without mtp at all.

Tensor split works with dense models, doesn't work with Flash Next (allocating 23087.12 MiB on device 0: cudaMalloc failed). It feels like llama treats it like a 24GB vram card, not 2x12. Anything I should adjust here? 

Is anyone running on a similar setup? If you could share your working llama recipe for it, I would appreciate it. 

  
Thank you
```



document fetched 2026-10-12 00:37 · canonical https://reddit.com/r/LocalLLaMA/comments/1x3jnaw/please_recommend_qwen38flashnext_llama_parameters
