ASUS ProArt PX13 posts strong Strix Halo benchmarks

This is a compelling “portable LLM box” data point: a thin-and-light laptop is pushing serious local inference numbers without resorting to a desktop GPU rig. The caveat is that the measurements are community-grade, but the performance-to-acoustics curve still looks unusually strong for something this small.

• –Performance mode peaks around 66-73W, 77C, and 47 dB, while still delivering roughly 1,007 tok/s prompt throughput and 69 tok/s generation on Qwen3.5
• –Balanced mode looks like the practical middle ground: materially lower noise and power, with only a moderate hit to throughput
• –Power saving mode is still usable for local inference, which matters more than peak numbers if you actually want to work on battery or keep fan noise down
• –The setup is especially interesting because it pairs Strix Halo with Linux and Vulkan/llama-bench, which is exactly the stack local-LLM enthusiasts care about
• –The post is missing Gemma4 results, so the comparison is incomplete, but the initial Qwen3.5 run already makes the point