DeepSeek V3 closes the year: open weights, frontier-adjacent results
Released on 26 December 2024, it is among the strongest openly available models yet. The year ends roughly where the plateau story said it would not.
On 26 December 2024 DeepSeek released V3, a large mixture-of-experts model with openly published weights. Its own evaluations put it ahead of other open models and close to leading closed ones on several standard tests.
A mixture-of-experts design activates only a fraction of its parameters for any given answer, which is how a very large model can be comparatively cheap to run. That architectural point is the reason the release matters beyond the leaderboard.
Three weeks, two lessons
In November the story was that scaling had plateaued. In December, o3 posted a large jump on a reasoning benchmark and an open model arrived within reach of the closed frontier. Both stories are true, and together they say something useful.
The frontier still moves, but the gap behind it is closing faster. Whatever the best closed model can do, expect something openly available to do most of it within months. For a buyer, that is the most consequential trend of the year.
Efficiency is where the competition actually is. Not raw capability. Cost per useful answer, and what hardware it takes.
The caution that belongs here
Open weights mean you may download and host the model. They do not mean the hosted service is a suitable place for your client data. Those are two completely different decisions, and the second depends on the operator, the jurisdiction and the contract, exactly as with any other provider.
Self-hosting, meanwhile, is not free. A model of this size is a serious infrastructure commitment, and for the overwhelming majority of SMEs a hosted service from a provider you have a contract with remains the right answer.
What to carry into January
Not a new model. A posture: assume capability will keep getting cheaper and more widely available, and stop treating your choice of provider as a long-term commitment.
Build so the model underneath can be swapped. Keep contracts short. Put your effort into the parts that do not become obsolete — clean data, a documented process, a named person who checks the output. That is the work that still pays in 2026.