Run a 30-billion-parameter LLM on four Raspberry Pi 5s for under $400—no GPU, no cloud bill
Need on-device AI but hate GPU prices? A hobbyist just squeezed Qwen3-30B-A3B onto four stock Raspberry Pi 5 boards and still hits 13 tokens/s. The trick is a mixture-of-experts model that swaps active parameters in and out of RAM, plus distributed-inference code that treats each Pi as a shard. Total cost: <$400 retail. Use it for private coding assistants, robots, or field deployments where cloud is impossible. Full build sheet and binaries are in the GitHub discussion.
hacker_news · 6 min read