Writing Lab
The smallest model that does the job
NVIDIA's own researchers argue that most of the work inside AI agents is better done by small models. For Swiss companies that is good news on cost, speed and privacy.

It is odd to hear the company that sells the biggest chips argue for smaller models. In June 2025 a team at NVIDIA Research published a paper with a blunt title: "Small Language Models are the Future of Agentic AI". Their point is simple. Most of what an AI agent does all day is narrow and repetitive. It reads a document, pulls out five fields, fills in a form, calls a tool, writes a short summary. That does not need a model trained to discuss Kant and write sonnets.
The authors call a model small if it fits on an ordinary consumer device and answers one user's requests quickly enough, which today means most models below 10bn parameters. The economics are striking. The authors estimate that serving a model of 7bn parameters is 10 to 30 times cheaper in latency, energy and computing than serving one of 70bn to 175bn. They looked at three open-source agent systems and estimated that small, specialised models could reliably handle between 40% and 70% of the calls those systems make to large models.
Frontier where it matters
None of this means large models are finished. When a task is open, new or genuinely hard, the frontier models from Anthropic, OpenAI and Google are still far ahead, and worth paying for. The NVIDIA team suggests mixing the two: small models for the routine, a large one for the exceptions. A well-run company works the same way. You do not send the finance director to post the invoices.
Small models stay at home
Small models have a second advantage that counts for more in Zürich than in San Francisco. They run on modest hardware, a single graphics card in a Swiss data centre or a server in the company's own rack. The data does not have to leave Switzerland, or even the building, which makes life easier under Swiss data protection law and with clients who ask where their files go.
Switzerland now has an open option of its own. In September 2025 EPFL, ETH Zürich and the Swiss National Supercomputing Centre released Apertus, a model in sizes of 8bn and 70bn parameters. It was trained on 15 trillion tokens in more than 1,000 languages, with about 40% of the data in languages other than English. It includes Swiss German and Romansh. Unlike most open models it publishes its training data and recipes as well as its weights, so a bank or a hospital can see what went into it. In July 2026 the same institutions released Apertus 1.5, which also reads images and audio and may be used commercially under the Apache 2.0 licence. With it came Apertus Mini, 16 compact models for machines with little computing power.
Fit the model to the job
Our rule at OptAI® is to start from the task and pick the smallest model that does it well enough, measured on the customer's own examples. Often that means fine-tuning a small open model for one job, such as reading supplier invoices or sorting tenant requests, and calling a frontier model only when the small one is unsure. The bill falls, the answers come faster and the data stays at home.
The race to build the biggest model will go on, and it is good that someone runs it. Most companies do not need to enter it. They need the smallest model that does the job, and someone who knows how to tell.
Sources
- Belcak et al., NVIDIA Research: Small Language Models are the Future of Agentic AI, arXiv, June 2025
- NVIDIA Technical Blog: How small language models are key to scalable agentic AI
- EPFL AI Center: Apertus, a fully open, transparent, multilingual language model, 2 September 2025
- Swiss AI Initiative: About Apertus
- EPFL AI Center: Apertus 1.5, building the next generation of open AI infrastructure, 24 July 2026
More writing
IntegrationThe plumbing is the productRivals adopted Anthropic's MCP. OpenAI and 19 investors put over $4bn into a firm that places engineers with customers. AI's value is in the connections.Author: Rea Stark
World modelsLanguage was the easy partYann LeCun's AMI Labs raised $1.03bn for world models. Fei-Fei Li and Google DeepMind work on them too. The next AI has to understand a kitchen.Author: Rea Stark
ZürichZürich has the labs. Now it needs the usersGoogle, OpenAI, Anthropic, Meta and Microsoft all do frontier AI research in Zürich. Only a third of Swiss SMEs build it into their work.Author: Rea Stark