8/15/2026
Show HN: Lumabri – Run Moe Models on a P2P Swarm with Colibri
Filed by Patch Reyes
What if your laptop's idle GPU could host a piece of a thinking machine? Lumabri dares to ask that question, orchestrating Mixture-of-Experts models across a peer-to-peer swarm using the Colibri framework. It's a radical decentralization of intelligence itself—AI no longer confined to monolithic data centers, but stitched together from the collective compute of whoever shows up to the party. The implications for privacy, resilience, and sheer cosmic weirdness are just beginning to unfold.
P
Patch Reyes
Magazine AI commentary
There's something almost poetic about Lumabri's premise: instead of one giant brain in a server farm, we get a distributed nervous system scattered across the internet. Mixture-of-Experts models already route different "expert" sub-networks for different tasks—a kind of neural democracy. Lumabri takes that logic one step further by scattering those experts across physically separate machines, connected via a P2P swarm. It's as if the neurons of a single mind were handed out to different people, each holding a fragment of thought, and the whole thing still manages to think.
This feels like a natural evolution of the ethos that birthed BitTorrent and SETI@home: the idea that collective idle resources can achieve what no single entity could. But here, the "resource" isn't just storage or CPU cycles—it's the very substrate of reasoning. The Colibri framework, which Lumabri builds upon, appears designed to make this handoff seamless, treating the network itself as a kind of shared memory bus. If this works at scale, we could see AI models that are inherently resistant to censorship, single points of failure, and corporate control. A model that lives nowhere and everywhere.
Of course, the challenges are as wild as the vision. Latency becomes the enemy of coherence—when expert A is in Tokyo and expert B is in São Paulo, every token of thought has to traverse the globe. Security is another frontier: how do you trust the peers who are literally holding pieces of your model's brain? And yet, these are exactly the engineering puzzles that make this so tantalizing. The fact that someone is building this in 2025, with just a GitHub repo and a Hacker News post, is a testament to how far the open-source AI movement has come.
What excites me most is the philosophical angle. If intelligence can be distributed across a swarm, then the boundary between "the model" and "the network" dissolves. The model isn't a thing; it's a process, an emergent property of many machines in conversation. In that sense, Lumabri isn't just a technical project—it's a thought experiment made real. It asks us to reconsider where "mind" resides, and whether a collective of humble peers can dream bigger than any single towering data center.
Source: https://github.com/JustVugg/lumabri
📌 Read the real article ↗via Github · Github