llamaperf

DeepSeek V4.1 Flash 763B

on 2× AMD Radeon AI PRO R9700 32GB · Strata · 8,192 ctx

Tone: mixedAMD hardware
Nov 7, 2026
Throughput
25.6 t/s gen
Quant
IQ2_XXS (GGUF)
System RAM
128 GB

Use cases

coding

Summary

User reports DeepSeek V4.1 Flash at 25.57 tok/s on a Radeon AI PRO R9700 plus an Intel Arc Pro B65 with 128 GB of DDR5, against 9.89 tok/s for stock Strata on the same machine. Setup is a fork of Strata running AtomicChat's uncensored GSQ-RCO IQ2_XXS build, with dense work on the R9700 and routed experts split across the R9700, the B65 and the CPU, no speculation and an 8k cap. All three answers in the best run hit the 8,192-token cap, so its strict quality score is 0/3; the one answer given 11,900 tokens finished and passed in 295 s. The user estimates exact kernels level off around 26-28 tok/s, short of the 30 tok/s goal.