Apple MLX adds RDMA over Thunderbolt to run Meta's 816B Muse Spark Open on two Mac Studios

Apple and Meta collaborated so that Meta's newest open-weight model, the 816B-parameter Muse Spark Open, runs locally on Apple hardware. The reference setup links two M5 Ultra Mac Studios, each configurable with up to 512GB of unified memory, and reaches 36.73 tokens per second. That is fast enough for interactive agentic coding.

The enabling work landed in MLX, Apple's open-source array framework. It added three things. Neural Accelerator support lets M5-generation GPUs use their dedicated matrix units. RDMA over Thunderbolt lets two or more Macs share tensors with low latency, effectively turning them into one inference node. New mixture-of-experts kernels efficiently route tokens through Muse Spark Open's sparse expert layers. Apple demonstrated agentic workflows such as bug fixing running end to end with no cloud calls.

The timing collides with NVIDIA's RTX Spark laptop launch the next day. Both companies are pitching unified memory as the path to serious local AI, but at different scales: a single 128GB laptop versus a multi-terabyte desk cluster. For enterprises with data-residency constraints, running an 800B-class model on roughly two workstations is a notable cost and governance alternative to cloud APIs.

There are caveats. Two maxed-out M5 Ultra Mac Studios cost far more than a typical developer workstation, so this is a prosumer or team setup. Meta's Muse Model License carries limited commercial terms, which the open-weights community contrasts with Apache 2.0 and MIT releases this week. And 36.73 tokens per second is single-stream throughput; concurrent-user serving will look very different.