Meta released Muse Spark Open on October 15, an 816-billion-parameter AI model that developers can run locally using two M5 Ultra Mac Studios. Meta collaborated with Apple to optimize the model for this setup, which requires 512GB of unified memory per machine.
Local deployment could allow companies to process sensitive information on their systems while reducing reliance on cloud inference services. However, the required hardware represents a substantial investment. Businesses must assess whether locally hosted AI agents can compete economically with cloud-based alternatives.
Muse Spark Open uses a mixture-of-experts architecture that activates approximately 42 billion of its 816 billion parameters for each token it processes. This reduces the computation required to generate responses, although the full model still needs considerable memory to store its weights.
Designed for coding and multi-step agent tasks, the model supports a one-million-token context window. It accepts text and images, plus audio and video, with text as its output. As an open-weight model, developers can download its trained parameters and run or modify it under the terms of Meta’s license.
How two Macs share the work
Apple’s unified-memory architecture allows the CPU and GPU to access the same memory pool, while its MLX framework distributes inference across connected Macs. Each machine stores a portion of the model and handles part of the computation. The machines exchange intermediate results as responses are generated.
This arrangement allows multiple Macs to run a model whose memory requirements exceed the capacity of a single machine. Apple tested a four-bit version of Muse Spark Open, which uses lower numerical precision to reduce memory consumption. Even with this optimization, the model occupies approximately 493GB before additional working memory is allocated.
Apple’s benchmarks recorded 24.99 tokens per second across two Macs and 36.73 across four, using an 8,192-token context and tensor parallelism. However, the tests generated only 128 tokens, so the results do not establish performance at the model’s full one-million-token context or during sustained agent workloads.
An agent that generates text quickly may still take considerable time to complete a coding task if it must inspect files or call external tools. Failed operations can require additional attempts. The reported generation speeds therefore provide an initial measure of local inference performance, rather than the time required to finish an entire workflow.
What the agents demonstrated
Meta’s local demonstrations included changes to Ice Cubes, an open-source iOS Mastodon client. According to Meta, the model fixed bugs and added features before completing a full build. Another demonstration involved DaVinci Resolve, where the agent reviewed footage and cropped selected moments into vertical video.
The demonstrations show Muse Spark Open performing coding and video-editing tasks, although the published examples do not establish how consistently it handles similar work across unfamiliar projects. Failed builds or incorrect edits could increase the time employees spend reviewing and correcting its output.
Meta’s model card acknowledges errors in multi-step reasoning and recommends safeguards for deployed agents. Organizations running the model locally would need to manage file permissions and require approval before sensitive or irreversible actions, particularly when agents can execute code or modify company records.
Running Muse Spark Open locally requires a substantial hardware investment
Apple’s M5 Ultra Mac Studio starts at $5,499, but the demonstrated configuration requires substantially more memory than the entry-level model. Apple says the 512GB version will become available in late October, which means Muse Spark Open’s release precedes the stated availability of the demonstrated hardware configuration.
According to The Wall Street Journal, the required Mac hardware costs upward of $40,000. Local inference eliminates per-token cloud API charges for workloads processed on those machines, but introduces an upfront purchase along with electricity and maintenance expenses.
To compare local inference with hosted services, businesses should distribute the hardware cost across its expected service life and account for ongoing expenses. Staff time spent administering the system or correcting agent errors also affects the cost per successfully completed task.
Companies with frequent AI workloads can spread the equipment expense across more tasks, while occasional users may find hosted services less expensive because they pay according to consumption. Hardware utilization therefore influences whether local inference can produce savings over time.
The comparison also depends on how much existing cloud usage the local installation can replace. If agents still depend on external models for certain tasks, businesses may continue paying API charges alongside their hardware expenses.
What the collaboration means for businesses
Local inference could give companies more options for processing sensitive information within their own computing environments. Prompts and model processing can remain on company-operated hardware, although agents connected to external services may still transmit data outside the organization.
For agents that access internal repositories or customer records, privacy depends on where information travels throughout the workflow. A locally hosted model may process those inputs privately, but integrations that send information to external APIs introduce separate data-handling requirements.
The collaboration could also increase demand for Apple’s high-memory Macs while giving Meta another way to distribute its open-weight models. These are potential commercial benefits of the partnership rather than confirmed sales or adoption results.
Businesses can evaluate Muse Spark Open before purchasing the demonstrated hardware. Meta makes the model’s weights available through Hugging Face and identifies hosted API providers, including Fireworks AI and Together AI.
However, the Muse Model License includes commercial restrictions. Certain model-as-a-service businesses exceeding $20 million in aggregate revenue must obtain a separate agreement with Meta, while internal use and specified other activities are exempt from that requirement. Organizations should review the license conditions before deployment.
Meta and Apple have demonstrated that an 816-billion-parameter model can run across desktop computers, although the required high-memory configuration is scheduled for late October and carries a substantial cost. The demonstration establishes technical feasibility, while sustained performance and simultaneous workloads remain insufficiently tested in the published results.
Companies with frequent AI workloads or sensitive data requirements may have reasons to consider local deployment. Whether the investment delivers savings will depend on how much cloud usage the hardware can replace and how consistently agents complete their assigned tasks. Lower hardware costs and more efficient models could make local deployment accessible to more businesses.
