GitHub - synchopate/llama-modchip: Run 284B DeepSeek-V4-Flash (<59GiB RAM, 44.4 tok/s pp, 9.1-14.7 tok/s) & 30B MoEs (25-111 tok/s) on a pure 64GB CPU at 256K ctx (0 GPUs). SwiGLU-H2, Stockfish-PVS L3 & 13ms Claude Code restore.

GitHub user synchopate released llama-modchip, a project that runs the 284B-parameter DeepSeek-V4-Flash model on a pure 64GB CPU with no GPUs, using under 59GiB of RAM at a 256K context window. The setup achieves 44.4 tokens per second in prompt processing and 9.1–14.7 tokens per second in generation, while 30B mixture-of-experts models reach 25–111 tokens per second. The project introduces SwiGLU-H2 and Stockfish-PVS L3 techniques and claims a 13ms Claude Code restore. It matters because it lets large MoE models run on commodity hardware without GPU investment, potentially widening access to frontier-scale inference. Next, watch for community benchmarks, upstream adoption, and whether the token rates hold across different CPU configurations.