# Should i try moving from Qwen 3.8 27B Q6_K swift 1.5 to QFN Swift 1.5 IQ3_XXS GSQ RCO? — retained record

> Retained reddit record from reddit:LocalLLaMA.

Canonical: https://highsignal.app/data/records/71eb36060cf0f785

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


source: [reddit](https://highsignal.app/data/reddit) (reddit:LocalLLaMA)  published 2026-10-13 15:53 ingested 2026-10-14 02:43

[original source ↗](https://www.reddit.com/r/LocalLLaMA/comments/1x50d7y/should_i_try_moving_from_qwen_38_27b_q6_k_swift/)

## retained text



```
I have my inference rig set up now for Qwen 3.8 27B Swift 1.5 Q6\_K at 262144ctx with KV at Q8/Q8.

My rig has 40gb of vram over two radeon VIIs (2x16gb) and my laptops Rtx3080 Max Q (8gb), using rocm and vulkan. Additionally there is 64gb of ddr 4 ram. The current setup runs at around 450tps PP16384 and 15-30tps output using dflash2.

But i am unsure whether i should try moving to the new Qwen 3.8 Flash next Swift 1.5 at a quant of only iq3\_xxs in terms of overall smartness - mostly for agentic development but also general agentic work in terms of design or "do task x for me".

Asking an ai model of choice to research this topic has not yielded any conclusive evidence for me - so heres my question:

Given my rock solid qwen 3.8 27b setup - do you think i should invest my time and tokens into getting a low Qwen 3.8 Flash next 1.5 GSQ RCO IQ3\_XXS quant to run at possibly Q8/Q8 KV cache? Do you see a chance of it being smarter than my current setup at comparable speeds?

I see quite some hurdles to be taken because with a mixed backend of vulkan and rocm i can forget the use of things like strata and will have to pursue my own fork of some sorts - probably llama.cpp with heavy modifications.
```



document fetched 2026-10-14 00:42 · canonical https://reddit.com/r/LocalLLaMA/comments/1x50d7y/should_i_try_moving_from_qwen_38_27b_q6_k_swift
