Until now, “bringing AI into the company” meant one of two things: paying a subscription for every employee, or paying per use through an API. Either way, your data travels to someone else’s servers and the bill arrives every month. On 10 August, Meta moved that boundary: it released Muse Glimmer, a 30-billion-parameter model under the Apache 2.0 licence that runs entirely on a single powerful computer. No cloud, no per-seat fee, no internet connection required.
It is not the first open model, and it will not be the last. But it is the clearest signal yet of where this is heading, and it deserves five minutes from anyone making decisions at an SMB. Here is what happened, what it actually means, and what you should do about it (spoiler: probably nothing drastic).
What exactly happened
According to Meta’s official announcement, Muse Glimmer is a dense 30B-parameter model that accepts text and images and is optimised for agent tasks: calling tools, writing code, organising files, retrying when something fails. Thanks to 4-bit compression, the model takes up under 20 GB and runs on machines with 24–32 GB of memory: a high-end MacBook or a PC with a consumer graphics card such as an RTX 5090. Uncompressed, it would need around 55 GB.
Two details matter more than the spec sheet:
- The licence is Apache 2.0. The same permissive licence behind half the business software you already use. You can download it, modify it and use it commercially without paying Meta or asking permission.
- It performs at the level of its class. In the published tests it beats comparable open models (Gemma4-31B, Qwen3.6-27B) on roughly half of a battery of two dozen benchmarks, with particular strengths in web research, code generation and chart analysis.
The model is already available on Hugging Face and works with easy-install tools such as Ollama and LM Studio.
What “open and local” means (and what it does not)
The jargon is worth translating, because this is where you find out whether it affects you:
- Your data never leaves your office. A local model processes invoices, contracts or case files without anything travelling to OpenAI’s, Microsoft’s or Meta’s servers. For work with sensitive data — accounting firms, law practices, clinics — that is a change of category, not of degree.
- The cost stops being monthly. No per-user fee, no usage bill: you pay once for the machine it runs on, and from then on the marginal cost of each query is electricity.
- It works without internet. Anecdotal for many, decisive for some.
And what it does not mean: it is not free (hardware capable of running it costs several thousand euros), it does not install itself (someone has to set it up, connect it to your systems and maintain it), and it is not at the level of the best cloud models. A local 30B is a competent professional; the top paid models are still the panel of experts.
The numbers, in practice
| Option | You pay | Your data | Where it fits |
|---|---|---|---|
| Subscription (ChatGPT, Copilot…) | Per user, per month | In the provider’s cloud | A personal assistant for each employee |
| Cloud API | Per use | In the provider’s cloud | Automations with variable volume |
| Local model (Muse Glimmer) | Hardware once + maintenance | In your office | High, steady volume with sensitive data |
The maths shifts with volume: if you process forty documents a month, the API wins hands down. If you process four hundred a day containing confidential information, a dedicated machine running with no per-use cost starts to make a lot of sense.
When this matters for your SMB
Three concrete situations where a local model like this changes the decision:
- You shelved an automation over privacy. If somewhere in a drawer there is a “we would love to automate this, but we are not uploading that data to the cloud”, that objection just lost most of its force. The whole process can run on a machine you own.
- Your sector demands provable confidentiality. In front of a client or a regulator, “the data never leaves our machines” is a much shorter sentence to defend than any data-processing annex.
- You have high, steady, repetitive volume. Sorting operational email, extracting data from documents, generating reports: agent tasks, exactly what this model is optimised for.
If any of those sound familiar, the move is not “install Muse Glimmer”: it is sizing the process — volume, hours, data sensitivity — and deciding with numbers whether it belongs in the cloud or on-premises. That analysis is exactly what we do in our AI and automation projects.
When it does not (most cases, today)
Let us be blunt, because that is why we write: if your use of AI is each person drafting, summarising and preparing proposals faster, a subscription remains unbeatable on cost and simplicity, as we covered a few days ago when comparing ChatGPT, Copilot and custom AI. Setting up a local server for that is buying a truck to fetch the bread.
Nor is this the moment to rush. Open models improve every few months and hardware keeps getting cheaper; whoever sets up today without a clear use case will have paid the early-adopter tax without the benefit.
What to do this week
Nothing spectacular: note the trend. AI capable of working with your documents no longer lives only in three providers’ clouds; it can also live on a computer in your office, under a licence that ties you to no one. That widens your options and strengthens your negotiating position, even if you never install a local model.
And if you have that sensitive-data process you have been meaning to automate for months, the objection holding it back may no longer exist. Tell us about it and we will tell you, with numbers and no strings attached, whether your case calls for local AI, a cloud API, or neither. The third answer is an honest one too.