Meta released Muse Spark Open’s weights on October 15, 2026, while also listing hosted API providers. Its model card describes a mixture-of-experts model with 816 billion total parameters, 42 billion active during inference and a one-million-token context window. Hosted access and running the weights on owned hardware are distinct routes. Apple’s same-day MLX test shows what one local Mac setup actually requires.
Why one 512GB Mac is insufficient
The 42-billion active-parameter figure describes how much of this mixture-of-experts model participates in computation at a time. It does not mean the complete model stores only 42 billion parameters. Apple estimates that the native BF16 weights alone would need more than 1.6TB of memory. Its 4-bit NVFP4 version reduces the model footprint to about 493GB, but inference also needs memory for intermediate work. Apple says the working model therefore cannot fit on a single Mac Studio with up to 512GB of unified memory.
Apple split the model across two M5 Ultra Mac Studios using MLX. In its pipeline configuration, one computer processes the first portion of the model’s layers and sends the intermediate result to the second. The test setup used two Thunderbolt cables between the machines and RDMA over Thunderbolt 5. Apple says this communication path depends on the driver introduced in macOS 26.2 and MLX version 0.30.1 or later. Those are prerequisites for Apple’s reported setup, not a claim that every deployment must use Macs.
What Apple measured
In Apple’s published decode test, an 8,192-token context followed by 128 generated tokens yielded 15.74 tokens per second across two Macs with pipeline parallelism. Two Macs using tensor parallelism, which distributes work within layers, reached 24.99 tokens per second. Four tensor-parallel Macs reached 36.73 tokens per second.
These are Apple’s measurements on specified hardware, precision and workloads. They measure generated-token throughput, not full request latency or model quality. The figures do not establish the same speed at the model card’s one-million-token maximum context, on other hardware or under a production load.
Weights, hosted APIs and operating terms
Meta lists downloadable BF16 and NVFP4 weights as well as third-party hosted APIs. Its model card’s deployment guidance says a conventional GPU setup needs at least 700GB of VRAM for NVFP4 or 2TB for BF16, plus an inference framework. Apple’s distributed MLX configuration is a separate route using multiple machines and unified memory. A hosted API avoids operating that local hardware but is not local inference.
The weights use Meta’s Muse Model License, which includes a usage policy, a “Built with Muse Spark Open” display requirement and additional terms for qualifying model-as-a-service businesses. Operators should check the terms against their intended use rather than treating “open weight” as an unrestricted deployment permission. Meta’s model card also warns that outputs can be wrong and recommends safeguards appropriate to applications that let the model use tools or take actions.
For a team considering local deployment, the decision is about the complete system: model precision, memory beyond the weights, inter-machine communication, workload-specific latency, licensing and application safeguards. Apple has demonstrated one workable configuration, not a one-box deployment or a general performance guarantee.
