Apple and Meta have jointly optimized Meta’s 816-billion-parameter Muse Spark Open model to run locally across two M5 Ultra Mac Studio systems. The setup requires 512GB of unified memory per Mac and a Thunderbolt 5 connection between the machines, putting a model of this scale within a desktop-class configuration rather than a conventional cloud deployment.
The model still needs substantial compression. In a 4-bit quantized form, it occupies about 493GB, down from more than 1.6TB in its native format. Additional memory is needed during inference, which is why a single 512GB Mac is not sufficient for this demonstration. Apple Silicon’s unified-memory architecture allows that memory to be shared with the GPU.
Using Apple’s MLX software, Apple reported the following response-generation performance with an 8,192-token context window:
- Two Mac Studio systems: about 25 tokens per second in the faster tested configuration.
- Four Mac Studio systems: about 37 tokens per second.
Tokens are units of text processed by the model, not a direct word count. The figures are Apple’s test results, and performance can change with longer documents or different workloads.
Meta has demonstrated the model handling local development work, including modifying an iOS app, fixing bugs and adding features. In another example, it used DaVinci Resolve to analyze video footage and select segments for vertical clips. Such workflows could be relevant for teams that need to process source code or internal files without sending them to a hosted AI service.
Muse Spark Open is downloadable as an open-weight model, although its use remains subject to Meta’s license, including commercial terms. The Apple collaboration does not make it Mac-exclusive; compatibility also extends to other hardware platforms:
- NVIDIA hardware
- AMD hardware
- Intel hardware
The two-Mac configuration provides a concrete local deployment path for a very large model, but its practical performance beyond Apple’s 8,192-token benchmark will depend on the context length and workload being run.
