APREX · EST. 2026
Not all open models are equally open
In the AI industry, "open source" is often used loosely. Some models release only the weights. Others release code but not training data. Some permit commercial use, but with heavy restrictions. The result is a landscape where "open" can mean anything from full reproducibility to pure marketing.
Ai2 is trying to shift that definition with Olmo 3 and the update Olmo 3.1. The institute describes not just a model, but a model pipeline: the entire lifecycle from data and pretraining through post-training, reinforcement learning, checkpoints, and evaluation.
That matters because genuine openness isn't just about downloading a file. It's about being able to understand how the model came to be, where it can be changed, and what kinds of decisions shaped its behavior.
Open weights let you run the model. An open model pipeline lets you understand and change it.
What Olmo 3.1 adds
Ai2 launched Olmo 3 in November 2025 and later updated it with Olmo 3.1. The update includes, among other things, Olmo 3.1 Think 32B and Olmo 3.1 Instruct 32B.
According to Ai2, the Think variant builds on an extended reinforcement learning run, with clear improvements on tasks such as AIME, ZebraLogic, IFEval, and IFBench. The Instruct variant is aimed at chat, tool use, and multi-turn dialogue.
The point isn't that Olmo suddenly beats every closed model. The point is that the development process is more transparent than that of most competitors.
Why this is research, not just a product
The Olmo series is of particular interest to universities, public-sector organizations, and research-adjacent companies. When data, code, and checkpoints are available, other groups can investigate how specific properties emerge.
That means more than benchmarking. If a model makes errors in legal language, specialized Norwegian terminology, or medical text, it is valuable to know whether the problem originates in the data, the tokenizer, pretraining, instruction tuning, or evaluation.
Closed models typically provide only the end result. Olmo tries to expose more of the causal chain.
Relevance for Norway
Norway has strong research communities but limited access to frontier-scale compute compared to the United States and China. Fully open models are therefore important. They make it possible to build on international research without starting from scratch.
For Norwegian language models, the public sector, and domain-specific AI, the Olmo approach may be more relevant than pure leaderboard winners. A model that can be adapted transparently, documented thoroughly, and evaluated locally may be better suited to sensitive or regulated domains.
This applies in particular to language, law, healthcare, and public administration, where trust and traceability matter more than raw demo impact.
Open weights are not enough
The industry has long celebrated "open weights" as a victory. And it is one. But open weights without training data and a recipe offer limited insight. You can run the model, but you know little about why it responds the way it does.
Ai2's argument is that open-source AI should be more like science: reproducible, verifiable, and possible to build upon.
That puts pressure on other players. If a model is marketed as open but provides no insight into its data or training, users should ask: Open for whom, and for what purpose?
The limitations
Olmo 3.1 is still not a magic answer to every need. Running and adapting 32B models requires expertise and infrastructure. Full openness can also make the system more complex to understand for ordinary product teams.
Furthermore, "fully open" is not the same as risk-free. Models can still hallucinate, produce incorrect code, reproduce biases, or be used misleadingly. Openness provides a better opportunity to find and fix problems, but no guarantee that the problems are gone.
Conclusion
Olmo 3.1 is one of the most important open-source releases because it shifts the question from "can I download the weights?" to "can I understand the model?"
For Norway, it is a useful reminder. If we want to build AI that can be used in the public sector, in research, and in regulated industries, it is not enough to choose the model that looks best in a demo. We need models that can be examined, documented, and adapted using open methods.
Olmo 3.1 shows what that can look like.