Lubuntu Beats Windows 11 in llama.cpp

Hot take: this is less a “Linux is always faster” story and more a reminder that local LLM performance depends heavily on which part of the stack is stressed; once llama.cpp leans on the CPU, Windows looks much worse here.

• –Fully GPU-offloaded runs show a steady Linux edge, usually around 4% to 8% in generation and low single to low double digits in prompt processing.
• –The biggest gap appears in hybrid runs (`-t 8 -tb 8 -fit on`), where Lubuntu roughly doubles prompt throughput versus Windows on Qwen3.5-35B-A3B and GPT-OSS-120B.
• –Gemma-4-E4B-it is the mildest case, which suggests the delta is workload-sensitive rather than a universal platform law.
• –The benchmark is useful but not definitive: driver versions, Windows build, compiler/toolchain differences, and prebuilt vs self-compiled binaries could all be contributing.
• –For people choosing an OS purely for local inference, the practical conclusion is that Linux seems worth considering if you run mixed CPU/GPU workloads or care about every last bit of throughput.