Research

May 27, 202614 min read

Shackled to the Machine: A DeepSeek-R1-8B Record on 4× Raspberry Pi 5

A dense 8B reasoning model sustains 8.32 tok/s bit-exact decode on four Raspberry Pi 5 boards, 29% above the prior four-node figure. The finding: dense decode is bandwidth-bound, so the firmware and the operating system move it; the compute kernels measure zero.

May 22, 202611 min read

Pushing Four Raspberry Pis to the Limit with distributed-llama

How we ran a 30B-parameter Mixture-of-Experts model (Qwen3-30B-A3B) on four CPU-only Raspberry Pi 5 boards at 15.143 tok/s decode — bit-exact and +16.1% over the public record — and hit the LPDDR4X memory wall.