Many firms hesitate to use language models because contracts, receipts or personnel data must not leave the building. Since April 2026 there is a serious alternative: Google released the fourth generation of its open model family under a free licence, in sizes from edge devices to workstation servers, with a very long context window and processing of text, image and audio.
Such a model runs on your own hardware or inside your own platform environment. Every data point therefore stays in-house, and questions about foreign access rights simply do not arise. Typical uses are extracting data from receipts, multilingual enquiries, and spotting anomalies in financial or production data.
Our assessment: for the highest quality the large vendor models remain ahead. For sensitive data the local variant is good enough today and it is an architectural decision, not a matter of belief.