One Radeon AI PRO R9700 (32 GB) + 125 GB RAM, Linux. Qwen3.8-Flash-Next IQ3_S on Strata, 0.2.6 vs 0.2.7, medians of 3 runs.
Prompt tok/s, before / now:
8K: 2,047 / 2,505 (+22%)
32K: 2,491 / 2,802 (+12%)
Decode tok/s, before / now:
8K: 82.9 / 94.6 (+14%)
32K: 64.9 / 85.1 (+31%)
Same card, just a software update: 32K decode up 31%, and 30 test tasks passed in every arm.
via Strata 0.2.7 release notes https://t.co/OESP7ifgcE · cc @AMD@RadeonPRO #LocalLLM