Really good stats! Sadly them being dense models will make them slow compared to something like Qwen3. Honestly my main reason not to use Mistral models when I need something on-prem with limited hardware (think Nvidia L4 GPUs).