# Qwen 3.8 Flash Next - 2 x R9700 vs. 3 x R9700 - 2 GPUs win — retained record

> Retained reddit record from reddit:LocalLLaMA.

Canonical: https://highsignal.app/data/records/561f34ca8a490408

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


source: [reddit](https://highsignal.app/data/reddit) (reddit:LocalLLaMA)  published 2026-09-04 16:42 ingested 2026-09-05 02:42

[original source ↗](https://www.reddit.com/r/LocalLLaMA/comments/1w79mpc/qwen_38_flash_next_2_x_r9700_vs_3_x_r9700_2_gpus/)

## retained text



```
My 3 x R9700 GPU system was running last weekend with Qwen 3.8 Flash Next, and after I saw various results here I wondered - could I get something similar with just 2 GPUs? The TLDR is yes, and with some optimization (not MTP yet) I was able to get even better token generation performance.

My hardware: X570 running x8 / x8 (with 3 GPUs x8 / x8 / x4 \[chipset\]), 64GB DDR4, Ryzen 9 5900XT, and R9700 32GB GPUs. Windows 11.

I ran a business writing test running AtomicChat AD-4.27bpw Q4\_K\_M quant. 

    .\llama-cli.exe `
      -m "Qwen3.8-Flash-Next-AD-4.27bpw-Q4_K_M-M64-00001-of-00033.gguf" `
      -f prompt.txt `
      -n 49152 `
      -c 131072 `
      --threads 12 `
      -b 8192 `
      -ub 2048 `
      -fa on `
      --parallel 1 `
      -ngl 999 `
      --jinja `
      --reasoning-format deepseek `
      --temp 0.7 `
      --top-k 20 `
      --top-p 0.95 `
      --min-p 0.0 `
      --reasoning off `
      --cache-type-k q8_0 `
      --cache-type-v q8_0

pp 242 t/s, generation 35 t/s. BETTER results overall than running 3 x R9700. the slow lane was hurting me, and Qwen doesn't need the extra VRAM. 

Net-net - Qwen 3.8 Flash Next is amazing
```



document fetched 2026-09-05 00:42 · canonical https://reddit.com/r/LocalLLaMA/comments/1w79mpc/qwen_38_flash_next_2_x_r9700_vs_3_x_r9700_2_gpus
