The Macs you already own, as one supercomputer.

Research. The transport substrate is published as Go modules: mlx-go-iroh ships peer discovery, gossip, and content-addressed blobs. Quorum-gated DiLoCo sync and peer-to-peer weight sharing run over it as live, untagged datapaths. In a small measured run, two machines fine-tuned, exchanged, and robust-merged real LoRA adapters to byte-identical files across three rounds. Coordinated training across the mesh is still being built. Modeled at $0 in any commercial planning.

• Memory-weighted scheduling — a 128GB Mac and a 48GB Mac pool their capacity automatically. Pipeline parallelism proportional to each node's unified memory.
• Content-addressed weight sharing — model weights move peer-to-peer over LAN, Apple AWDL, and a global iroh mesh, verified against Hugging Face SHA-256 digests rather than trusting any peer.
• iroh transport — QUIC links and signed gossip via mlx-go-iroh. LAN and gossip by default; global discovery is opt-in.
• Quorum-gated sync — DiLoCo-style updates apply only when a quorum of distinct peers publishes the same signed digest. Fail-closed: the live datapath stays off unless explicitly enabled.

mlx-mesh is R&D. The substrate is real and published: mlx-go-iroh (peer discovery, gossip, content-addressed blobs) is a tagged Go module, and weight sharing plus a quorum-gated DiLoCo sync loop run as live datapaths on top of it. What remains is the hard part: feeding verified deltas into a full training loop, cross-device latency, and heterogeneous-compute consistency. Production use is at your own risk; commercial planning treats it as upside, not revenue.

mlx-mesh sits alongside mlx-go as the distributed-compute arm of the Apple Silicon work. When a model doesn't fit on one Mac — or when a regulated team wants training workloads to stay inside the building — mlx-mesh turns idle devices on the LAN into coordinated capacity. cove is the isolation companion when those training workloads need sandboxing.