APREX · EST. 2026
The model that changed the mood
DeepSeek-R1 was one of the most dramatic AI releases of 2025. Not because it had the slickest demo, but because it struck a nerve: what if advanced reasoning isn't something only the largest American labs can deliver behind closed APIs?
DeepSeek published R1 with open weights, a GitHub repository, and an MIT licence. That gave researchers and developers the ability to download the model, test it locally, distil smaller variants, and build their own systems on top of it.
What R1 actually did
DeepSeek-R1 became known for reasoning: mathematics, code, logic, and multi-step problem solving. That puts it in the same category that made OpenAI's o1 famous — but here the model was released in a format that could be studied and reused.
DeepSeek also published distilled models based on Qwen and Llama variants. That made the technology accessible to far more people than those who can run the full model.
DeepSeek-R1 turned reasoning into an open-source competition, not just a closed premium feature.
The MIT licence matters
The MIT licence makes commercial use and further development easier than many alternative model licences. For startups, consultancies, and research environments, that means less legal friction during the pilot phase.
It does not mean every question is resolved. Training data, safety, geopolitics, and censorship still need to be assessed. But the licence lowers the threshold for practical experimentation.
Norwegian relevance
For Norwegian organisations, R1 is less about switching everything to DeepSeek and more about stress-testing assumptions. Does advanced reasoning have to be purchased as an expensive API? Can smaller distilled models handle internal tasks well enough? Is local deployment realistic for bounded domains?
This is particularly relevant in education, research, the public sector, and industry, where language, data flows, and cost must be kept under control.
The risks must be taken seriously
DeepSeek is a Chinese company, and that alone means many Norwegian organisations need to carry out additional due diligence. This applies especially when using a hosted API. Running open weights locally presents a different risk profile than sending data to an external service — but it does not eliminate the need for security testing.
Models can also produce wrong answers with great confidence. Reasoning traces often look convincing even when they are incorrect.
Conclusion
DeepSeek-R1 was more than a model release. It forced the industry to acknowledge that open AI can be competitive even on reasoning.
For Norway, the most important takeaway is pragmatic: build your own benchmarks, test open alternatives, and consider local deployment where data and cost matter more than maximum demo impact.