A 10 year old Xeon is all you need (for 26B-A4B MTP Drafters without GPU)

A developer successfully ran a quantized Gemma 4 language model (26B parameters) on a decade-old Intel Xeon server with 128GB DDR3 RAM and no GPU by leveraging advanced optimization techniques in ik_llama.cpp, particularly speculative decoding with a smaller draft model. The key to making this work on memory-constrained hardware was understanding and manually configuring dozens of obscure optimization flags that standard tools like Ollama don't expose, addressing the "memory wall" bottleneck that dominates LLM inference on older systems.