DeepSeek-V4.1-Flash: new encoder-decoder architecture, native vision, and V4 Pro gets rerouted from September 14

Mistral released Large 4 as a public preview today. It is a 1.05-trillion-parameter Mixture-of-Experts model with 49B active parameters, text and vision input and a one-million-token context window. Mistral trained it from scratch on 3,800 Nvidia Grace Blackwell GPUs in its own European datacenters and runs it there, end to end, under European law. The weights are due as open weights by the end of October.

It doesn't quite match the leading US models, but if you need your data and your provider to stay in Europe, it is the first serious option at this size.

At DevDay on 29 September 2026, OpenAI announced more than 20 new features across ChatGPT, Codex and the API. The common thread: agents that no longer just answer a question but take on ongoing work, operate computers and react to events in other applications – and, if you want, run entirely inside AWS.

With taktoro.de and taktoro.com, our new product for companies in Germany is live: time tracking that checks every entry against the German Working Hours Act (ArbZG) before it is saved. Breaks, rest periods, maximum working hours and the record-keeping duty are built in, not booked as add-ons. Developed and hosted in Hamburg.