APREX · EST. 2026
The reasoning shock got a follow-up
DeepSeek-R1 was one of the model releases that genuinely rattled the AI market in 2025. Not because it was the most polished package, but because it demonstrated that strong reasoning models could come from somewhere other than the largest American laboratories.
With DeepSeek-R1-0528, DeepSeek updated the model and made it available on Hugging Face under the MIT licence. That matters because R1 was never just a model. It was an argument: advanced reasoning can be trained and shared in ways that reshape the economics of the entire field.
DeepSeek turned reasoning into a question of price, not just prestige.
What R1-0528 represents
R1-0528 is an update to DeepSeek-R1, published on 28 May 2025. The full model is available on Hugging Face, and DeepSeek has also released a distilled 8B variant based on Qwen3.
That last point matters. Large reasoning models are demanding to run, but distilled variants make the technique more accessible to developers, researchers, and smaller companies that do not have frontier infrastructure at their disposal.
The RL recipe behind the breakthrough
The DeepSeek-R1 paper describes how reasoning capabilities can be strengthened through reinforcement learning. R1-Zero was particularly striking because reasoning behaviour emerged without traditional supervised fine-tuning first, before DeepSeek built on that foundation with a more usable R1 model.
This makes the R1 series significant for the research community. It does not merely showcase a finished product, but a training direction: models that learn to apply more explicit reasoning to mathematics, code, and complex tasks.
Why Norwegian communities should care
Norway is not going to win the AI race by training the largest general-purpose models from scratch. But Norwegian research communities, startups, and organisations can gain an edge by understanding, testing, and adapting open reasoning models.
R1-0528 is relevant for mathematics, code, technical analysis, and internal decision-support tools where it matters that the model works more thoroughly than a quick chatbot. At the same time, Norwegian users must test language quality, safety, source usage, and political or cultural biases themselves.
Openness does not mean absence of risk
DeepSeek's models have also raised questions about censorship, jurisdiction, data security, and evaluation practices. Saying that a model is open is not enough. Organisations need to know where it is run, what data it is exposed to, what constraints it operates under, and how it behaves on their own tasks.
For Norwegian organisations, R1-0528 should therefore be treated as a research and pilot tool first. It may prove highly useful, but it should not be uncritically introduced into sensitive decision-making pipelines.
Conclusion
DeepSeek-R1-0528 demonstrates that R1 was not a one-off event. Reasoning models continue to push the market, both technically and economically.
For Norwegian teams, the best response is neither hype nor fear. It is targeted testing: your own tasks, your own data, your own security requirements. DeepSeek has made it harder to pretend that only the largest closed models matter.