Running a 235 billion-parameter AI model used to require renting a small fortune’s worth of cloud compute. Mesh LLM wants you to do it with the GPUs already sitting in your house, your neighbor’s house, and maybe a few strangers on the internet.
The open-source project, led by developer Michael Neale and endorsed by Jack Dorsey, pools idle Nvidia GPUs across multiple machines into a peer-to-peer network capable of distributed inference on large language models.
How it actually works
Mesh LLM creates a decentralized mesh of GPUs spanning macOS, Linux, and Windows machines, with a heavy emphasis on Nvidia’s CUDA architecture. Instead of routing your AI queries to OpenAI’s or Google’s data centers, the system keeps everything local or distributed across connected nodes with zero central server involvement.
The project exposes an OpenAI-compatible API, which means developers can swap in Mesh as a drop-in replacement for cloud-hosted models without rewriting their applications. Under the hood, a novel pipeline approach called “Skippy” partitions massive models across multiple nodes. A 235 billion-parameter Mixture of Experts model, for instance, gets sliced up and spread across whatever hardware is available on the network.
Data transmission relies on the iroh P2P library, which handles NAT traversal, the networking headache that makes peer-to-peer connections difficult when devices sit behind routers and firewalls.
On a LAN setup, Mesh LLM achieves roughly 16 tokens per second on that 235B model. Over wide area networks, speeds range between 10 and 25 tokens per second.
The project supports dozens of open-weight model families, including Qwen, Llama, GLM, and DeepSeek variants.
The rewards layer
A related project called Meshtrain, launched on August 27, 2026, adds an incentive structure for GPU providers. Contributors who share their compute resources earn “MeshCoin” through a P2P local ledger.
That mainstream application is Block’s Buzz app, where Mesh has been integrated for community compute sharing. Block, the fintech company run by Dorsey, has been steadily expanding its footprint in both Bitcoin and decentralized infrastructure.
Privacy as a feature, not an afterthought
One of Mesh LLM’s core selling points is its zero-central-server architecture. When you run a model through the network, your data never touches an external data center.
Performance does take a hit on wide area networks compared to local setups. The architecture works best when nodes are geographically close and connected by low-latency links.