llamaperf

Qwen3.8 125B (6B active) Flash-Next

on Intel Arc Pro B70 · Strata · 262,144 ctx

Intel hardware
Nov 20, 2026
Throughput
51.1 t/s gen · 1456.2 tokens/s prompt processing (prefill)Prompt processing is how fast the input is read, not how long you wait for the first word. How to measure time to first token.
Quant
IQ3_XXS (GGUF)
System RAM
48 GB
VRAM reported
32 GB

Summary

User reports Qwen3.8-Flash-Next at 51.09 t/s generation and 1456.24 t/s prefill on a single Intel Arc Pro B70, with a 4096-token prompt. Setup is Strata with IQ3_XXS weights, a 262144-token context, MTP and drafting off, and a 20000 MiB expert cache with 32768 KV tokens resident. The launch uses all 32GB of the card's VRAM. A sweep from 4096 to 250000 prompt tokens gives generation between 43.61 and 51.09 t/s and prefill between 1053.93 and 2250.96 t/s. Peak prefill was 2529 t/s. The host is a Ryzen 7 5700X3D with 48GB RAM, with a B580 for display.