By Roman Bodnarchuk
10XAI.news | Special Series 2 of 14 | Research checked October 18, 2026
A desk photograph can make a major computing shift look ordinary.
Two computers. A screen. A place to work.
The useful question for a business owner is whether that setup changes which work can stay inside the organization.
What the primary sources establish
On October 15, Meta released Muse Spark Open, an open-weight model with 816 billion total parameters and 42 billion active parameters. Meta describes running it on two M5 Ultra Mac Studios and releases it under a license with limited commercial terms. [1]
Apple’s accompanying technical report is more specific: the model’s parameters require more than 1.6 TB in BF16; the evaluated NVFP4 version has a footprint of about 493 GB before intermediate computation needs. Apple says it does not fit on a single machine in that setup. [2]
My judgment
The opportunity deserves attention. It also deserves a specification sheet.
Owning two Mac Studios does not establish that you own the configuration used in a published demonstration. The chip generation, memory capacity, model precision, and connection between machines all matter.
Before buying anything, write down the task. A document classification workflow and a coding assistant have different requirements. A quiet overnight batch job and an interactive assistant have different tolerances for delay.
I would assess a local deployment in this order:
Fit: can the chosen model and workload fit within available memory?
Quality: does it meet your standard on representative work?
Responsiveness: how long do you wait before the first useful output?
Operations: who maintains it, updates it, and investigates failures?
Control: where do inputs, logs, backups, and tool calls actually go?
Local inference can be one component of a private workflow. It does not establish that every connected service is local, or that access controls are correct.
Your 10X action
Inventory the hardware you already own. Select one permitted workload to evaluate before considering an upgrade. Use a model suited to that hardware and approved for your use.
Copy this prompt
Create a feasibility checklist for local AI using my actual hardware specifications: [details]. My workload is [task]. Identify missing measurements, memory and context requirements, external dependencies, and a representative quality test. Separate published benchmark configurations from what has been verified on my machines. Do not recommend a purchase without identifying the constraint it would solve.
Reply with the task you would most like to keep inside your own environment.
Roman Bodnarchuk
10XAI.news