richardslatter/r820-llm-public is a repository holding a case study and tooling for a modified ik_llama.cpp that runs DeepSeek V4.1 Flash on a 2012 Dell PowerEdge R820. Its subject is running a 552-billion-parameter mixture-of-experts (MoE) model on an old server without adding any new parts. The repository includes the patch plus scripts for building, model conversion, and serving. No official release version was given.
The main changes are a vector instruction patch that works on chips from the 2011 server generation without a strong extended instruction set, and a NUMA layout that lets each core read its own memory. The GPU handles half of the dense computation and most of the prefill. The speeds recorded in the README are 9.95 tokens per second for decoding and 101 tokens per second for prefill. The README also explains that applying the feature that leverages model-generated drafts in the code pushed the recorded figure up to 12–13 tokens per second.
The README provides a staged performance table for AVX1 optimizations and the division of work between memory and GPU, and also discloses the optimization attempts that had no effect. This repository is therefore best read as a technical record of the modifications and trial and error involved in running a model on an old server. No license information is given, so the reuse terms cannot be confirmed from the repository alone.