arXiv preprintarXiv:2611.06238
Zuse: A Decision Model for European Languages
Umair Mehmood Imam, Bineric AI and OsloMet, Norway
Zuse is the first family of decision models with published results beyond English: 4B, 9B and 27B models that read a document and return a calibrated probability for each allowed answer instead of writing text. Measured in 27 European languages on 11 sources the models never trained on, the 27B reaches 87.5% accuracy overall and 94% on banking intents, and the paper sets out the three findings on calibration, automation and reach that make it work.
Zuse 27B accuracy by task
On sources the model never trained on, across 27 languages
- Banking94%
- Reading93%
- Sentiment90%
- Topics90%
- European languages
- 27
- held-out sources
- 11
- questions in the suite
- 20,044
- languages tested for reach
- 210
Abstract
Decision models read a document and return a probability for each of a fixed set of options instead of generating text. Several appeared in 2026, but none has published results outside English. We describe Zuse, three decision models (4B, 9B and 27B) that train a low-rank adapter and a pointer head on frozen open-weight bases, targeting 27 European languages, and use them to study how this design behaves across languages and tasks. On a broad European suite of 11 sources excluded from training, covering 9 task groups and 27 languages, a development set that also guided model selection, the 27B model reaches 87.5% accuracy with a pooled expected calibration error of 0.016, and 0.035 when weighted per source. Post-hoc analyses of these single-seed results give three findings. First, miscalibration is structured by task rather than language: on parallel reading and topic items in 26 languages the confidence gap varies about as much as sampling noise, while across sources it spans −0.07 to +0.09, with knowledge tasks overconfident and acceptability judgements underconfident at all three sizes, so one fitted temperature cannot correct both. Second, the share of decisions that can be automated at a 5% error budget follows chance-corrected skill (Spearman 0.97 over 33 source and model pairs) more closely than raw accuracy (0.91). Third, reach beyond the training languages is set by the base model: across 210 languages, changing the training run moved scores by at most about 2 points within a size, while languages outside the base’s documented list trail by 23 to 28 points at every size. On English public suites Zuse is behind Jev and Kev, including Kev models on the same bases. We also report two pitfalls in building multi-source training data: label-word collision between topic schemes, and an injection augmentation that taught a shortcut. Zuse can be tried at https://bineric.com/zuse.
Cite this paper
@misc{imam2026zuse,
title={Zuse: A Decision Model for European Languages},
author={Umair Mehmood Imam},
year={2026},
eprint={2611.06238},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2611.06238},
}