Lesson learned. Same Strix Halo. Same Qwen3.8 but OS ... Windows 11 → 37.94 sec Ubuntu → 25.27 sec That’s 33% less time just by changing the software/OS stack. Hardware Busters retested the same BOSGAME M5: 🧠 Ryzen AI Max+ 395 💾 128GB unified memory 🎮 96GB graphics allocation On Ubuntu 26.04.1, the long Qwen3.8 response finished in 🪟 Windows → 37.94 sec 🐧 Ubuntu → 25.27 sec And the first text appeared about 11 seconds sooner. The Ubuntu setup used llama.cpp with Mesa/RADV Vulkan and ROCm paths. This is a pretty big deal for anyone buying these 128GB Strix Halo boxes for Local AI. Before spending money on new hardware, keep in mind the OS/runtime stack; it may be leaving a lot of performance on the table. 🔗 Great article from Hardware Busters in ALT. Link in ALT