There is a list in most small businesses that nobody has written down. It is the work you have quietly decided not to hand to an AI tool: the client file, the medical intake, the payroll spreadsheet, the contract still under negotiation. Not because AI could not help, but because you were not comfortable uploading it to someone else’s server.
The short version: Meta has released a free, open-weight model called Muse Glimmer that runs offline on a single consumer GPU. For anyone weighing AI for small business, it does not make AI cheaper, and for most owners it changes nothing at all. It changes exactly one thing: the work on that unwritten list can now be done by a capable model that never sends a byte outside your office.
What did Meta actually release?
Meta Superintelligence Labs published Muse Glimmer on August 10, a 30-billion-parameter model distilled from its larger Muse Spark system and released under the permissive Apache 2.0 license. The weights are on Hugging Face, free to download, and they run through tools that contractors and technical staff already use: Ollama, LM Studio, llama.cpp, and vLLM.
The interesting number is not the parameter count, it is the memory. At full precision the model needs over 55 GB of video memory, which is server hardware. Compressed to roughly 4-bit precision, Meta reports it fits in about 32 GB with 0.2% average quality degradation, or about 24 GB with 1.0% degradation. Those are consumer numbers. A high-end desktop graphics card or a well-specified Mac clears that bar.
It was built for agent work rather than conversation: tool calling, multi-step tasks, and recovering when a step fails partway through. It reads images as well as text, handles over 100 languages, and holds a context window above 131,000 tokens. On agentic benchmarks Meta places it ahead of similar-sized rivals, scoring 75.5 on MCP Atlas against 54.2 to 62.5 for Gemma4-31B and Qwen3.6-27B. It also loses some of those comparisons; Qwen leads on computer-use and terminal tasks. As always the benchmark table belongs to the vendor, and the questions worth asking about any model claim have not changed.
Can you really run AI for small business on your own computer?
Yes, and at this quality level that is genuinely new. When Moonshot AI published the largest open-weight model ever a few weeks ago, our honest read was that you would never actually download it, because running it required a datacenter. The free license was real and the access was theoretical.
Muse Glimmer inverts that. The hardware fits under a desk. Speed is workable too: Meta reports 233.4 tokens per second on an RTX 5090 with its speculative decoding enabled, up from 74.9 without it, and roughly 50 tokens per second on an Apple M5 Max. That is faster than you can read.
Which makes this the first time the honest answer to “could I run my own AI?” is yes. It is not yet the first time the honest answer to “should I?” is yes, and the gap between those two questions is where the money gets lost.
What does running AI locally actually cost?
Free weights are not a free system. Three costs arrive with them.
The hardware is the obvious one. A machine with 24 to 32 GB of video memory is a deliberate purchase, not a laptop you already own. Set that against a per-seat AI subscription for a small team, and the payback period on a dedicated machine is long. It only shortens if the thing runs constantly.
The second cost is a person. Somebody has to install it, connect it to whatever it is meant to act on, and notice when it quietly stops working. A subscription includes that person in the price. Self-hosting does not, and this is where most small deployments actually die.
The third cost is the one people miss. A local model is frozen. Muse Glimmer’s knowledge stops on January 4, 2026, and it will not learn anything after that date unless somebody downloads a replacement and sets it up again. The hosted tools you pay for get quietly better while you sleep. This one does not. You are trading continuous improvement for control, which is a reasonable trade only if you actually need the control.
Which work belongs on hardware you own?
Here is the exercise, and it takes about ten minutes. Do not start with your budget. Start with that unwritten list.
Write down every task you have refused to give an AI tool because of what the data is, not because of what the tool costs. Signed contracts. Client financials. Anything covered by a confidentiality clause you personally signed. Patient, client, or student records. The employee complaint you had to document. Board material before it is public.
That list is the entire case for local AI. Not the pricing, not the open-source principle, not the enthusiasm of whoever forwards you this article. If the list carries real weight, then a machine in your office that reads those documents and never phones home is worth the hardware and the setup, because the honest alternative was never a cheaper tool. It was continuing to do that work entirely by hand.
If the list is empty or trivial, you have your answer too, and it is that today’s news is not for you. Keep paying for the hosted tool. That is not a failure of ambition, it is an accurate read of your own risk.
One caveat worth holding onto: local does not mean safe by itself. A model running on your own hardware can still be pointed at accounts and files it should never touch, and that failure looks much the same as it does in the cloud. The permissions question stays exactly where it was. What changes here is who holds the data, not who checks the work.
Should you switch your business to local AI?
Almost certainly not, and it is worth saying plainly, because open-weight releases attract a lot of noise. Meta is making a policy argument here as much as a product one, with Zuckerberg publicly pressing Washington to clear the way for open-weight AI. That argument is about American competitiveness and regulatory capture. It is not about your quoting process.
What today actually hands an owner is narrower than a switch, and more durable. It is an option that did not exist last week: a serious, capable model you can run where your data already lives, at zero cost for the software itself, for the specific work that was never going to leave the building anyway. Most owners will look at that option and decline it, correctly. The point is that you now get to make that call yourself, rather than having it made for you by what the vendors are willing to host.
Frequently Asked Questions
What is Muse Glimmer?
Muse Glimmer is a 30-billion-parameter AI model released by Meta Superintelligence Labs on August 10, 2026 under the open Apache 2.0 license. It is built for agent work such as tool calling and multi-step tasks, it reads both images and text across more than 100 languages, and it is designed to run offline on a single consumer GPU rather than in the cloud.
What hardware do I need to run AI locally?
Meta reports that Muse Glimmer needs over 55 GB of video memory at full precision, but compresses to roughly 32 GB with 0.2% quality degradation or roughly 24 GB with 1.0% degradation. In practice that means a desktop with a high-end graphics card or a well-specified Mac, which is a deliberate hardware purchase rather than a typical office laptop.
Is local AI cheaper than a subscription?
Usually not. The model weights are free, but you pay for a machine with enough video memory, for someone to install and maintain it, and for the fact that a local model is frozen at its training cutoff while hosted tools keep improving. Local AI wins on data control, not on price.
Should a small business run its own AI model?
Only if it has work that genuinely cannot leave its own network, such as signed contracts, client financials, or records covered by confidentiality rules. If that list of restricted work is short or empty, a hosted subscription remains the better choice for most small businesses.
So what is actually on your list of work you have refused to hand to an AI tool, and would a machine that never phones home be enough to change your mind about it?
