Small Models, Own Data: How SMEs Can Stay Independent with AI

Quotations, customer data, internal processes: do you really want to upload these to a third-party cloud in order to use AI? Small language models show that there is an alternative. They run locally, are economical to operate and can be adapted to specific needs, allowing SMEs to retain control over their data.

The most efficient model for a given purpose

ChatGPT has impressively demonstrated what AI can do, yet many companies face a dilemma: while the technology is impressive, scepticism about data protection is growing. An industrial company, for example, wants to automate cost calculations and customer projects, or a fiduciary firm uses AI to analyse documents. Both are aware, however, that trusting external cloud providers can be risky if highly sensitive trade secrets have to leave their own infrastructure.

Whereas large language models act as digital all-rounders, Small Language Models (SLMs) [1] behave more like precise specialists. They have considerably fewer parameters and are therefore optimised to solve specific tasks highly efficiently, either in training or while running as AI tools in combination with specific data. A key advantage is the hardware requirement: SLMs fit on an ordinary server or even a powerful laptop. An analogy helps to illustrate this: for the daily transport of smaller goods, one does not need a huge truck; a small delivery van is often far more efficient and practical.

Self-determination, in AI too

For companies, digital sovereignty means above all self-determination: it is about controlling precisely where their own data is stored, who has access to it and which providers they depend on technically. For SMEs, this is not abstract geopolitics but a pragmatic necessity. Those who lose control find their room for manoeuvre slipping out of their hands when external cloud providers change their data protection terms, raise prices or alter licences. Sovereignty thus protects a company’s own business model from external dependency. In view of possible forthcoming regulation, such as the EU AI Act, using local models also offers the decisive advantage of compliance: those who retain full control over their AI systems can meet regulatory requirements considerably more easily and proactively.

Potential and limitations of small models

A key advantage of SLMs is local operation: if data never leaves the company, the protection of highly sensitive information remains assured – and the AI works reliably even without an internet connection. Compared with huge models, SLMs are considerably cheaper to run and consume less energy, as they are optimised for specific tasks. Another plus is transparency: since many modern models are available as open source or under open licences – such as Mistral, Gemma or the Swiss model Apertus [2] – companies do not have to trust a «black box» blindly. It is possible to see exactly which version is active and, in part, what the training data was, which forms the basis for responsible and controllable use of the technology.

A small model can be trained on, or connected to, a company’s own technical terminology, specific documents or internal processes. In an industrial company, for example, an AI tool can be optimised to answer technical warranty enquiries, while a fiduciary firm uses an SLM to summarise complex reports consistently according to its own templates. In customer support or in the analysis of inspection records, too, SLMs thus enable highly specialised and reliable AI-supported processes.

SLMs are not all-rounders: for complex or abstract tasks, large models remain superior. Moreover, running SLMs requires a certain level of technical know-how for set-up and maintenance – a hurdle that poses challenges for many SMEs in day-to-day operations.

Conclusion

The largest model is not necessarily the best; the best is the one most suited to the specific purpose. While giants may impress, small models offer a decisive advantage: they enable SMEs to make full use of the enormous possibilities of artificial intelligence without giving up control over their data or their digital sovereignty.

Tomyra AI [3], a spin-off of Bern University of Applied Sciences, supports SMEs and other organisations in taking this step. Tomyra AI is an AI environment for organisations that work with client data, case files or trade secrets. The model runs in an isolated environment, in Switzerland or directly at the organisation. The AI comes to the data rather than the other way round. Organisations are also supported in designing their AI-based processes so that they deliver the greatest benefit, and in building the necessary skills in their teams.

References

[1] Further reading on the topic: https://learning.oreilly.com/library/view/hands-on-small-language/0642572259877/

[2] https://www.apertus-ai.org/

[3] https://tomyra.ai