A new project, llama-modchip, demonstrates running a 284B DeepSeek-V4-Flash model and 30B mixture-of-experts models entirely on a 64GB CPU with no GPU. It reports under 59GiB RAM use, 256K context, and throughput of 9.1–14.7 tok/s for the 284B model and 25–111 tok/s for the 30B MoEs.