agi

Why we included this project

If you have ever wanted to train or fine-tune a model across machines you and your team actually own, without spinning up a central cluster or renting cloud GPUs, this project is worth studying. It runs a peer-to-peer network where each node trains locally and shares compressed weight updates instead of raw data, using a DiLoCo-style scheme that squeezes each round's payload from megabytes down to tens of kilobytes. Consumer laptops, small VMs, and home workstations can pool into one mesh and run a collaborative training round together; the maintainers report 32 anonymous nodes training a language model in 24 hours this way. A separate pods feature lets a small group bundle their machines into a shared cluster for distributed inference. This is experimental research software, a proof of concept for decentralized training rather than a production platform, but for anyone curious about federated approaches that need no trusted infrastructure, it is one of the more concrete implementations to study.