There’s a server in my basement that has no business running a modern language model. It’s a repurposed HP Store Virtual storage box, roughly thirteen‑year‑old, two Ivy Bridge Xeons, no GPU. As of this week it runs Google’s Gemma 4, a 26‑billion‑parameter open‑weights mixture‑of‑experts model, at about 5 tokens/sec.

Running Gemma 4 on such antiquated silicon is hard because the optimized ik_llama.cpp fork assumes AVX2 and FMA3, instruction sets that didn’t ship until Haswell. The older Xeon E5‑2690 v2 (Ivy Bridge) only has AVX1, so the fast kernels fail to compile and execute.

Claude guided the patching: the hot paths were rewritten to fall back to scalar/SSE math when AVX2 is absent, and the graph builder was updated to split fused gates into separate mat‑mul calls. The resulting diff lives in the ik_llama.cpp branch #2138.

The final build produces text at pointer reading speed, roughly 5 tokens/sec on thirteen‑year‑old silicon, for about $300 of hardware. It shows that a legacy enterprise box can host a large, modern MoE model, providing a local fallback when paid APIs are unavailable or cost‑prohibitive.