XeStrata test raises AI text generation from 10.9 to 31.1 tokens per second

PC Watch measured 31.1 tokens per second with the XeStrata inference engine versus 10.9 with Strata on the same Intel Arc B580 computer, Notebookcheck reported. Both configurations ran the 125-billion-parameter Qwen3.8-Flash-Next model with the same settings. The computer paired a 12GB graphics card with 64GB of system memory. The selected compressed model configuration requires approximately 50GB of system RAM, so the result does not mean the model fits in graphics memory alone. The comparison measures text-generation throughput, without establishing equivalent gains in prompt processing, latency or accuracy. The reviewer also considered the B580 insufficient for heavier agent workloads, and the tested XeStrata package is no longer the latest release.