Back to News

Software Development · Oct 16, 2026

Can you run AI locally? Meta's Muse Spark Open and the case for open-weight models

On October 15, Meta released Muse Spark Open, an open-weight version of its Muse Spark model. According to Meta's research blog, it has 816 billion total parameters, with 42 billion active at any time, and a context window of 1 million tokens. Meta says it can run on two Mac Studio computers with M5 Ultra chips.

Mark Zuckerberg announced it on X. Alex Finn, who posts frequently about AI agents on X, shared his own setup running it locally.

For most of the past few years, "run a top-tier AI model in your own office" was not a realistic sentence. It is getting closer.

What open-weight means

An open-weight model is one where the company publishes the trained model itself, so you can download it and run it on your own hardware or your own cloud account. That is different from models like GPT-6 or Claude, which you can only use through the provider's service.

Open weights do not always mean open for any use. Meta released Muse Spark Open under its Muse Model License, which includes limits on commercial use. Read the license before building a product on it.

Why a business might want to run AI locally

Data never leaves your building. For businesses with strict confidentiality needs, such as legal, healthcare, or anyone with contracts that limit where data goes, this can be the deciding factor.

Predictable cost. You pay for hardware and electricity, not per token. For very high, steady volume, that can make budgeting simpler.

No surprise changes. A model you run yourself does not get updated, retired, or repriced without your say. That stability is valuable for software that needs to behave the same way for years.

Works offline. Useful for locations with unreliable internet or for systems that should not depend on an outside service.

Why most small businesses should not, yet

Running AI locally is real infrastructure. Someone has to set up the hardware, keep the software updated, monitor it, and secure it. A large model also needs serious hardware, and that hardware sits idle when nobody is using it.

For most small and mid-size businesses, a business plan from a major provider with strong data terms, like zero data retention, gives most of the privacy benefit with none of the maintenance. Hosted models are also usually ahead on capability.

A middle path

There is an option in between. Smaller open-weight models can run on modest hardware and handle narrow, high-volume tasks well: classifying documents, extracting fields, or summarizing internal notes. You can use a local model for sensitive, routine work and a hosted model for harder tasks.

That kind of hybrid setup is more work to build, but it can make sense for businesses with specific privacy or cost needs.

What this means for your business

  • Running capable AI on your own hardware is now realistic, but it is still an infrastructure project.
  • Check license terms on open-weight models before using them commercially.
  • Local AI makes the most sense for strict confidentiality, steady high volume, or long-term stability needs.
  • For most businesses, a hosted business plan with strong data terms is the simpler choice.

If you are weighing local versus hosted AI, or a mix of both, our software development and Managed IT teams can help you think through the tradeoffs and build what fits.

Sources