Mesh LLM pools strangers' GPUs for AI — the speed number was missing (iroh.computer)
Mesh LLM, built on the peer-to-peer networking library iroh, splits large AI models across whatever GPUs happen to be online: a laptop holds layers 0 through 15, a server holds the rest, and the pool answers as one OpenAI-compatible API. The project's own writeup shipped without a single benchmark.
A commenter dug the only number out of the model catalog anyway, a modest one for a flagship-sized model, right as four existing rivals for the same idea surfaced in the replies.