Local AI: what October's open-weight models change for an SME's data

In a single week, three "open" AI models made the news: Muse Spark Open from Meta on 15 October, Nemotron 3.5 Super VL from NVIDIA the same day, after Apertus 1.5, the Swiss model from ETH Zurich and EPFL, released in late July.

Behind the announcements, a very practical question arises for a Swiss SME: should you run AI on your own machines rather than sending your data to an online service? We propose a decision grid.

What has just been released, in a few facts

Muse Spark Open has 816 billion parameters, 42 billion of them active, and a context window of one million tokens. Meta presents it as the most capable open-weight model in the United States, according to its announcement of 15 October 2026.

Its size is its first obstacle. According to TechCrunch, running it locally requires at least 1 TB of memory, typically a Mac Studio with an M5 Ultra chip. That is not an ordinary office machine.

NVIDIA's Nemotron 3.5 Super VL plays in another league: 120 billion parameters, 12 billion active, and one specialty, documents. OCR, tables, multi-page PDFs, video. Its Bolt variants run on a PC with 24 GB of memory.

Apertus 1.5, released on 24 July 2026 by ETH Zurich, EPFL and CSCS, is under the Apache 2.0 licence, so it can be used commercially without special conditions. It now understands images and audio, and comes in compact versions, the Apertus Mini suite.

Three models, three philosophies: raw power, document specialisation and Swiss transparency. None of the three is a gadget.

Why the question arises now for an SME

Until now, local AI was reserved for well-equipped technical teams. Two things have changed: the quality of open models has caught up with online services on many everyday tasks, and the tools to run them have become commonplace.

Ollama, LM Studio or vLLM install a model in a few commands. NVIDIA explicitly names these tools for its Bolt models. A competent Swiss provider can now set this up in a day.

The economics follow. An API call costs little at low volume, but a workflow that handles thousands of documents a month changes the equation, especially when every document carries personal data.

The second driver is regulatory. The revised Swiss Federal Act on Data Protection has applied since September 2023, and in February 2025 the Federal Council confirmed a sector-specific approach to AI: no horizontal law, but existing obligations that apply in full.

For a fiduciary, a medical practice or an engineering firm, "where does my data go?" is no longer an IT question. It is a question of management responsibility.

Local or online service: the decision grid

We use four criteria with our clients. None is sufficient on its own, it is their combination that decides.

• Data sensitivity: trade secrets, health data, files covered by professional secrecy. The higher it is, the more local deployment weighs in the balance.

• Volume and regularity: daily, heavy use pays off a machine, occasional use does not.

• Performance needs: drafting, summarising and document extraction work very well locally, complex reasoning is often still better online.

• Available skills: a local model needs someone to update it, monitor it and secure it.

There is a third path, often the right one: an open model hosted by a Swiss or European provider, under contract. You keep contractual control of the data without managing the hardware.

Watch the licences. Muse Spark Open requires a "Built with Muse Spark Open" notice in any product that uses it, and derivative models inherit its conditions. Apertus, under Apache 2.0, has no such constraint. Have the licence read before any deployment involving clients.

Example: a Geneva fiduciary and its accounting documents

Take a fiduciary with 25 employees in Geneva, processing around 4,000 accounting documents a month for its clients: supplier invoices, expense claims, bank statements.

Today, data entry and checking take close to 120 hours a month. An online AI service could do the job, but the documents contain client names, salaries and bank details. Management refused to send anything to a service outside Switzerland.

The Ramp case is instructive. According to NVIDIA, this US fintech adapted Nemotron 3.5 Super VL to receipt extraction with 10,000 examples and a single training pass, reaching the level of the best models on the market for that task.

For our fiduciary, the reasonable scenario is a Nemotron Bolt or Apertus Mini type model, installed on a server on its premises or with a Swiss host, connected to its accounting software. Indicative budget: 8,000 to 15,000 francs for hardware or hosting in the first year, plus integration.

With 70% of documents processed automatically, around 80 hours freed up per month, the return on investment lands at about twelve months. And the data never leaves the company's perimeter.

What we take away

October's open models do not make online AI obsolete. They give SMEs a real choice, and a real choice forces you to think before you buy.

Three pieces of advice to move forward:

• Classify your use cases by data sensitivity before even talking about models.

• Test an open model on a repetitive, document-based task: that is where local deployment shines.

• Have the licence and the hosting contract reviewed before putting a single client in the loop.

Want to know where your company stands? Our AI Barometer gives you a reading of your maturity in ten minutes, data included. And to discuss it, contact us.