Explore distributed inference and training with MLX

A Podkey summary of Explore distributed inference and training with MLX, from WWDC 2026.Apple quietly put together something pretty interesting here: a way for multiple Macs to behave much more like a real high-speed compute cluster, using Thunderbolt 5 as the link. The big picture is RDMA over Thunderbolt, JACCL on top of that, and MLX to make it usable for actual inference and fine-tuning. And the practical hook is hard to miss. This can push model work across several Apple Silicon machines, including models too large for one box, while keeping everything local and private.RDMA over Thunderbolt 5JACCL turns the link into something usableMesh and ring without the tuning headacheTwo ways to split a modelWhat the speedups actually looked likeThe trillion-parameter angleFine-tuning scales tooSetup, APIs, and who this is forWhy this stands outThis podcast was created with Podkey. Make your own at https://podkey.fm