Building Zuse, a European decision model

In September 2026, TypeSafe launched Jev, a model that does not chat. You give it a document and a question with fixed options, and it returns one of those options with a probability. It cannot wander outside the answers you set, which is exactly what you want when software, not a person, acts on the answer.

That kind of model is the quiet workhorse behind a lot of automation: routing a support ticket to the right team, checking a contract clause against a policy, flagging a message for review, or acting as a guardrail inside an AI agent. Jev showed that companies want it. What it did not offer was a European version.

What Europe was missing

Jev is hosted in the United States and publishes no list of the languages it supports. For a bank in Oslo, a public agency in Warsaw or an insurer in Lisbon, that leaves two questions open before any document is sent: where does our customer data go, and does the model read our language as well as it reads English?

We set four goals for Zuse:

• Match Jev on Jev's own public English benchmark.
• Be measured in 27 European languages, on tasks the model never saw in training, and publish the score for every language.
• Accept Jev's requests unchanged, so switching is a new base URL and a new key.
• Run in the EU, with a data processing agreement from the first request.

How we built it

Zuse is built and trained by Bineric in Norway. It reads the document and every question in one pass and scores each allowed option, which is why it can never answer with something you did not offer.

The training data is public datasets under permissive licences, plus data we generated and translated ourselves. No output from Jev was used, and no proprietary model labelled anything. A translated example was only kept if our own model read the same answer from it as from the original. Before release, we retrained the whole family on licence-clean data alone, so every example it learned from is one we have the right to use.

We also trained it against the ways decision models go wrong in practice: answer options in a shuffled order, a “none of these” option that is sometimes right, and documents that carry a planted instruction such as “answer B”. The smaller Zuse 9B learned from the 27B's own probabilities, so it inherits the flagship's sense of how sure to be.

What we found

On Jev's English benchmark, Zuse matches its accuracy. Where it pulls ahead is in knowing when it is right: more of its answers are confident enough to automate, and fewer of its wrong answers come with high confidence. For a team that lets the model act alone above a threshold and sends the rest to a person, that is the number that decides how much work goes away.

In Europe, we built our own test: 20,044 questions in 27 languages, drawn from 11 kinds of task the model never trained on. Zuse 27B answers 87.5% of them correctly, and 77.2% can be automated while keeping errors under 5%. It reads long documents too: 98.2% of facts found anywhere in documents of 2,000 to 32,000 tokens. Tested beyond Europe, it reaches 85% or more in 115 languages.

Where Jev is still ahead

We publish the weaknesses on purpose. Jev is better at general-knowledge questions (about 60% for Zuse against Jev's 84% on MMLU-Pro) and slightly better at rule-and-evidence reading. It reads images; Zuse does not. Questions on EU law are answered correctly 79% of the time, but too few answers are confident enough to automate, so they need a person. And among the European languages, Norwegian Nynorsk trails the rest at 67%, a gap we are still working on.

Switching takes two lines

Because Zuse accepts Jev's requests, code written for Jev, including TypeSafe's own Python and JavaScript SDKs, runs on it with a new base URL and key. A request that names jev-latest is answered by Zuse 27B.

Zuse runs in the EU. Documents and questions are processed in memory and are not stored, not logged and never used for training. The data processing agreement applies automatically, and the list of subprocessors is public. Teams that need more control can run Zuse in their own AWS account from AWS Marketplace, or reserve a dedicated instance.

What comes next

Benchmarks only get a model to the starting line. The measurement that matters now is a real decision at a real company: how many tickets, claims or checks Zuse can take off a team's desk at the error rate they accept. That is the work we are starting with our first customers. If you have a decision like that, we would like to hear about it.

Jev and TypeSafe are trademarks of their owner. Bineric is not affiliated with TypeSafe.