Gemma 4 26B MoE runs at 5 tokens/sec on 2013 Xeon hardware

The Gemma 4 26B-A4B mixture-of-experts model achieves 5.2 tokens/sec decode and 16 tokens/sec prompt evaluation on dual Intel Xeon E5-2690 v2 processors. This setup uses Q8_0 quantization on legacy hardware lacking AVX2, FMA3, or dedicated GPU acceleration.