A hands-on test published in late August 2026 examines how far language models on local hardware have come. The verdict: they trail the large cloud services by roughly six months. A usable setup starts at around two thousand euros, with the graphics card as the main cost. Reducing model weights to four bits cuts memory requirements from over fifty to about thirteen gigabytes with no noticeable loss in quality. These models are strong at operating tools and weak at recalling facts.
When it pays off
Three arguments carry weight in the Swiss context: clarity about where personal data is processed, predictable costs at consistently high volume, and independence from a single vendor's model policy.
The sensible approach is rarely all or nothing, but a split by data class. Confidential material stays in house, while demanding and uncritical work runs in the cloud. The ERP system provides the structured foundation.