What if the AI model you want to run is bigger than any machine you own? Slice it into pieces.
Today me (Jair) and my friend Tiago are launching PieceOfMind (PoM), an open-source project that lets you combine GPUs from different computers, and different people, to run large AI models over the internet. In this video we run DeepSeek V4 Flash (~120GB) split across two home machines in different parts of the city, with only ~5ms of latency between them.
Most people think latency would make this unusable. After months of trial and error, we found it doesn’t. We call the technique LLM Pipelines. Unlike tensor parallelism, which needs the full model on every machine, POM slices the model by layers. Each GPU runs only the slice that fits in it, and only a small vector per token travels over the network, not the model weights. When a model doesn’t fit your GPU, this can be faster than offloading to RAM
