Qwen3.8-Flash-Next 177B running at 11–15 tok/s on a single RTX 5070 12GB + 32GB RAM DDR4

Benchmarking an LLM here with a NVIDIA RTX 5070 12 GB VRAM here

I had been working on a llama.cpp based expert streaming setup for Qwen3.8-Flash-Next 177B (UD-IQ3_XXS) on Windows. Benchmark is about 11.5 tok/s, up from roughly 7 tok/s on the inherited setup. In normal conversations I’ve seen 14–15 tok/s, and a long coding prompt generated 4,892 tokens at 10.15 tok/s and produced a working single-file Snake game.

Hardware: RTX 5070 12GB
32GB DDR4-2400
Ryzen 5 5600GT PCIe Gen3 Windows

The main gains came from fixing Windows I/O queue-depth issues, using one file handle per worker, and building a page-locked hot-expert tier so the GPU can pull hot expert weights more efficiently.

(In the video its around 16 minutes for 10k tokens and 10.41 tok/s

Output is quality gated against the control model and the published benchmark uses a heat file built from a separate prompt set.

Demos:

https://www.youtube.com/watch?v=cOPumMlyj_4

https://www.youtube.com/watch?v=rc-uTjVpXM8

In the GitHub I have things I've tried that didn't work and benchmark scripts, and methodology. If you guys have suggestions especially for streaming please let me know