A 10 year old Xeon is all you need | Hacker News

A Hacker News post showed Gemma 4 running locally on recycled server hardware: a single Xeon E5-2620 v4 with 128GB RAM and no GPU.

Using llama.cpp-based tooling and CPU-only inference, the setup delivered roughly reading-speed performance on a modern 26B model.

Commenters debated benchmarks, power draw, and how far older machines can now go with quantization and optimized runtimes.

The example suggests practical local AI is no longer limited to high-end GPUs, and even aging servers may handle useful LLM workloads.