Short answer: "AI that runs on your own machine" was the best-sounding headline of the week, but for an SME it is almost certainly not the answer to your data protection questions. Meta's new open-weight model really can be downloaded, and it really does run on a Mac. Just not the kind of Mac most companies own. The rest of the week's news was less glamorous, but it affects your decisions more directly: pricing, switching providers, and who checks AI companies from the outside.
Meta's Muse Spark Open: open, but not for your laptop
Meta released Muse Spark Open, an open-weight model, on 15 October. According to the Hugging Face introduction, it has about 816 billion parameters and a context window of one million tokens, it accepts text, images, audio and video, and it was trained mainly for coding and agentic tasks. Meta calls it the most capable open-weight model from the US to date. That is Meta's claim; there is no independent comparison behind it yet.
TechCrunch reports that the model was designed to run locally on Apple computers. The full model, however, needs at least 1 TB of memory, and it runs best on a Mac Studio with the M5 Ultra chip. In practice that means several high-end Mac Studios linked together. Hugging Face's own test shows that a heavily compressed version fits on a single 512 GB machine, using about 304 GB of memory and generating 18 tokens per second. It works. For daily work, it is slow.
What does this mean for an SME? Open-weight models matter: they create competition and give software vendors new options. But in a company of 30 to 100 people, the idea of "let's run it in-house so the data never leaves" rarely holds up. The hardware is only the start: someone has to operate it, update it and secure it. For most SMEs, the data protection risk is handled by a well-chosen business plan (the business tier of Claude, ChatGPT, Microsoft Copilot or Gemini), its contractual data processing terms, and a clear internal policy. Running your own model makes sense when, for one narrow, well-defined task, the data genuinely cannot leave the company. Even then, usually a smaller model.
What else matters
- Gemini 4 Argon: Google introduced its new frontier model on 13 October, which Google says is built for complex, multi-step work such as coding, finance and legal tasks. The introductory API price is $2 per million input tokens and $10 per million output tokens; after 30 November it becomes $4 and $20. If you build on it, budget with the doubled price.
- Replit switched its default model: according to TechCrunch, the developer platform's agent now uses OpenAI models by default instead of Anthropic's Claude. The CEO named cost, partnership and performance as the reasons, and Claude remains available as an option. The lesson is not which model is better. It is that a model switch is a business decision, and your system has to be able to handle it.
- Independent evaluators at Anthropic: on 15 October Anthropic announced that six staff members of METR, a nonprofit, now have "employee-like" access, roughly on par with its own risk assessment teams. According to Anthropic, it cannot veto their publications; it can only require redaction of predefined categories such as customer data. This is worth weighing when you choose a provider: who checks, from the outside, the company you trust with your data?
What to do now
- Before you plan your own AI server, write down exactly which data must never reach an external provider. The list is often shorter than you think. Our article on company data and AI tools helps with this.
- Review your AI subscriptions: are your colleagues on business or personal plans, and what are the data terms of each? The comparison is here: Claude, ChatGPT, Copilot or Gemini for an SME?
- If you build on an API, budget with the final price, not the introductory one, and check what happens to the cost if usage doubles.
- Plan for a model switch: keep your prompts and tests in a form that lets you try them quickly on another provider's model.
If you want help deciding which AI solution and which data setup fit your company, take a look at our AI solutions service.
