News · 19 November 2026

Local AI: language models in-house.

Anyone using AI usually sends their data to a cloud service. It does not have to be that way. Open language models have become capable enough to handle many tasks on your own hardware: in the server room, on a workstation or right inside a device. For companies with sensitive data this is becoming a real alternative.

Compact AI server with two graphics cards in a server room

Why local models are interesting now

Two developments come together. Freely available models have improved markedly and cover tasks such as summarising, classifying, extracting data or answering questions from your own documents well. At the same time the hardware required has become smaller and cheaper. A model that once needed a data centre now runs on a single server with a graphics card.

What speaks for running it in-house

Data sovereignty. Customer data, contracts, design documents or patient data do not leave the company. That simplifies compliance with Swiss FADP and GDPR and answers many questions before the legal department asks them.

Predictable cost. Instead of paying per request, you pay for hardware, power and operations. With a high, steady volume that pays off. With a few requests a day, the cloud remains cheaper.

Independence. The model does not change overnight, there is no dependence on a provider’s pricing and product decisions, and the solution works without an internet connection.

Where the limits are

The largest cloud models remain superior for demanding tasks, such as long multi-step analyses or complex software development. Running locally also means responsibility: procuring hardware, updating models, securing access and monitoring the quality of the answers.

In practice a hybrid is therefore often the sensible choice. Sensitive data and routine tasks stay local; for particularly hard tasks without delicate data, a cloud model is added.

AI right inside the device

Edge AI goes one step further: the model runs not in the server room but on the electronics inside the device itself. A camera recognises objects, a machine reports deviations or a device understands speech without data leaving the device. Here the interplay of hardware, firmware and model determines response time, power consumption and unit cost.

Support from inaffect

Your AI, your data.

inaffect AG supports you in this area: we clarify which tasks can be solved locally, select the model and hardware, integrate the solution into your systems and operate it on request. For edge AI we develop electronics, firmware and software from one team, compliant with Swiss FADP and GDPR.

Our AI services

We look forward to your inquiry.

One field is enough – we get back to you within 24 hours.