Meta’s New AI Can Run Locally. Should Your Business Try It?

The possibility is real. The business case is yours to prove.

Imagine a repair shop where the same three people are constantly interrupted: Where is the current intake form? Which document explains the inspection fee? What should a customer bring when dropping off a machine? The answers exist, scattered across folders and old files. An internal assistant that finds the right page could be useful.

On October 15, Meta released Muse Spark Open, an open-weight AI model. That means developers can obtain the model’s learned parameters and run it on suitable infrastructure, subject to its license. Meta also lists hosted providers for businesses that prefer to use someone else’s infrastructure. [1] (opens in a new tab) [2] (opens in a new tab)

For an owner, the useful change is another choice about where an AI system runs and who operates it. A local installation deserves consideration when keeping a particular workflow on controlled equipment solves a concrete problem—and someone can maintain it. For occasional questions, buying a collection of powerful computers is a substantial commitment before the assistant has earned its place.

Downloading a model also does not create a finished staff tool. Someone still has to connect approved documents, build the interface, manage access, and decide what happens when the answer is missing. That work determines whether this imagined shop gets fewer interruptions or another system to supervise.

“Runs on a Mac” needs a little translation

Meta’s hardware guide lists an Apple configuration of two M5 Ultra Mac Studios with 512GB of memory each, or four with 256GB each. These are specific reference setups with fast connections between machines. They should not be read as a claim that the demonstrated version will run on an ordinary office laptop. [3] (opens in a new tab)

There is a timing wrinkle. Apple’s August announcement says the 512GB configuration is coming in late October. As of this October 17 source check, that announcement does not establish that the two-machine setup is generally deliverable today. The four-machine configuration is another documented route; confirm actual stock, configuration and delivery with a supplier. [3] (opens in a new tab) [4] (opens in a new tab)

Apple lists a U.S. starting price of $5,499 for an M5 Ultra Mac Studio. That is a base price, not a quote for a 512GB computer or a complete AI installation. Memory upgrades, multiple machines, setup and support need their own estimate. [4] (opens in a new tab)

The reason for the scale is straightforward: the full model still has to be held in memory even though only part of it does the work for each generated token. Meta also warns that runtime overhead and the working conversation consume additional memory. These documented configurations are useful starting points for a specialist’s estimate, rather than a shopping list for every business. [2] (opens in a new tab)

Apple’s October 15 MLX article documents tests on connected Macs. Those results demonstrate a particular technical setup. They do not establish how accurately an assistant will interpret your shop’s documents, or how quickly it will serve several staff members at once. [5] (opens in a new tab)

Start with the question a colleague keeps asking

For the hypothetical shop, start with intake administration. Give a prototype a small, current collection: the approved drop-off guide, inspection-fee policy and contact directory. Leave repair diagnosis, safety instructions, payments and customer records outside this first version.

A staff member asks, “What do I tell someone bringing in an item for inspection?” A useful response would point to the applicable document, show its revision date, and offer a short draft the employee checks. If two policies disagree, the interface should flag the conflict and offer the source pages instead of confidently picking one.

The interface matters as much as the model choice: readable answers, keyboard-operable controls, visible source links and an easy route back to the original document. Staff should be able to keep working if the assistant is unavailable.

There is also a simpler competitor: a clearly named folder with good search. If that gets people to the right policy quickly, stop there. If questions regularly require combining several approved documents, a limited AI pilot becomes easier to justify.

Local control comes with local responsibilities

Running the model locally can avoid sending that inference request to an outside model provider. It does not, by itself, establish where every part of the application sends data. A builder still needs to inspect document storage, logging, backups, analytics and any cloud connections. This is an architectural distinction, not a privacy certification.

In the shop example, permissions should decide which documents can be retrieved before an answer is generated. Asking the model to keep a confidential folder secret is not a replacement for preventing access. A local machine also needs controlled accounts, updates and a recovery plan; someone must own those jobs.

Hosted access offers a way to explore the model without purchasing that hardware, using material appropriate for the provider’s terms. On October 17, OpenRouter lists $1.25 per million input tokens and $4.25 per million output tokens for Muse Spark Open. Tokens are the pieces of content the service processes and produces; these rates are usage prices, not the complete cost of an application. The listing also says EU in-region routing is not currently available. [7] (opens in a new tab)

Compare the whole cost of each route: building the assistant, preparing documents, checking answers, operating it and helping staff. For local deployment, add equipment and maintenance; for hosted deployment, add usage charges and the chosen provider’s conditions. A free model download does not erase those expenses.

The download has conditions, too. Meta uses its custom Muse Model License, including an attribution requirement and a usage policy. Have the person deploying the system check the applicable terms; “open-weight” is not permission to ignore them. [6] (opens in a new tab)

A small trial should be allowed to say no

Before choosing hardware, collect twenty representative staff questions using non-sensitive material. Include ordinary questions, one where the policy changed, one with conflicting documents and several that the supplied files cannot answer. Write down the correct source or the expected “not found” response before running the trial.

Compare ordinary search with the proposed assistant. Record time to a checked answer, wrong answers, missing source references, time spent correcting drafts and total operating effort. Measure elapsed time through human verification, not just how quickly text appears. These are proposed evaluation steps, not results from a test we performed.

Set the acceptance rules in advance. For this limited administrative pilot, a made-up fee or an answer based on a superseded policy should block expansion until investigated. A refusal to guess when the relevant document is absent is useful behavior. Keep the existing document path available throughout.

Only after the workflow proves useful should you compare local and hosted operation on the same tasks and required controls. A smaller model may be sufficient, or the original search tool may win. Muse Spark Open gives a business another option; the best first purchase may still be a few hours spent organizing the information its people need.