Meta released Muse Spark Open on October 15 with Apple collaboration aimed at local execution on high-end Mac desktops, according to CNBC. The report describes an 816-billion-parameter mixture-of-experts model and configurations involving two to four Mac Studios. This is a multi-machine deployment proposal, rather than a claim that an ordinary laptop can run the model.
Engadget reports a roughly one-terabyte memory requirement, with two 512GB M5 Ultra Mac Studios or four 256GB machines among the described arrangements. Its article says the 512GB version is expected in late October, without a specific ship date or price. Memory capacity and immediate hardware availability are therefore separate questions.
CNBC says Meta describes the release as open-weight with limited commercial use. It also reports that the companies claim local execution can improve security and save money. Those are company claims, not a conclusion from testing by this publication. The model’s availability should not be read as unrestricted commercial permission or as proof that a local installation will cost less for every team.
Engadget describes company benchmarks suggesting advantages over GLM 5.3 and Kimi K3 in agentic tasks and visual reasoning. A vendor comparison can help identify a model worth testing; it does not establish performance on a buyer’s own prompts, software or hardware configuration.
CNBC also reports that Apple published technical documentation about its MLX framework’s ability to split models across multiple machines. The companies declined to say whether the work marked the beginning of a deeper relationship. Those details keep the announcement bounded: the reported collaboration concerns this local-model deployment, while its longer-term scope remains unspecified. The report does not supply a complete configuration procedure.
Our assessment: the interesting change is the proposed deployment unit. A team considering the release would be assessing several computers as a working system. Merely adding their memory capacities together would not demonstrate that its own installation runs correctly. The runtime, model format and connection between machines would all belong in the acceptance test. We are identifying evaluation requirements, not claiming that a specific interconnect or configuration has passed our testing.
For example, a hypothetical software team could take a small set of representative maintenance tasks and compare the proposed local setup with its current service. It should record whether each task completes correctly as well as how long it takes. A model loading successfully would be an initial milestone. It would not establish that the system produces useful changes, handles tools correctly or remains responsive enough for the team’s workflow.
Keeping computation on owned machines would also leave practical controls to review. In a proposed pilot, we would check which applications and tools the installation can access, what it logs and whether any part of the workflow still contacts external services. Those questions do not assert a defect in Muse Spark Open. They distinguish the location of model execution from evidence about the behavior of a complete agent system.
There is a documentation limitation in our review. The public model repository was reachable, but our text extraction returned template material. We therefore do not present that extraction as a verified model-card review. The hardware and release details above are attributed to the two retrieved reports; license and runtime requirements need direct inspection before deployment.
The release is a concrete option to investigate for organizations interested in large local models. Its practical value will depend on a tested configuration, permitted use and results on the organization’s actual work. The available reporting establishes the announcement and described deployment approach. It does not settle those operational decisions for a prospective user.
Before planning a Muse Spark Open deployment, obtain the actual model license and supported runtime instructions, confirm the required Mac memory configurations are available, and test a representative workload before buying hardware.
