DeepSeek ships again, and the release cadence tells you something
V3.1 appeared on 21 August. Ten months into this blog, the interesting fact is no longer the model. It is how routine this has become.
DeepSeek released V3.1 on 21 August 2025, its latest open-weight model, following V3 in December and R1 in January. Alongside Qwen, Kimi, Llama and now gpt-oss, the count of credible downloadable models has passed the point where listing them is useful.
In January, a DeepSeek release moved stock markets. In August it is a Thursday.
What actually changed in eight months
Not much about the models, which improve steadily and predictably. What changed is the assumption underneath.
In January, the question a Swiss SME asked was whether a cheap open model could be trusted. In August, the question is which of eight roughly equivalent options fits, and the differences between them matter less than the difference between having a process and not having one.
The consequence for your decisions
Model choice has become a smaller decision than it was, and this blog has been saying so since March. Three things now matter more:
Where the processing happens, which is a contract question. Whether you can pin a version, which August made vivid. Whether your people actually use the thing, which 14 August covered and which is where most value is lost.
If you are still holding a decision open because a better model might arrive, note that a better model will always be arriving. That is now a permanent condition rather than a reason to wait.
The one thing worth watching
Not capability — availability in Europe. A model that is excellent but that your provider cannot host in a European data centre is not an option for a company with data-residency requirements, however good the benchmarks.
That is the question to ask about any model a vendor proposes this autumn, and it is a shorter conversation than comparing capabilities.