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.
Figures & tables
Figure 1: Zuse architecture: document and question in, one probability per option out. The low-rank adapter sits inside the frozen base’s layers.
Model
Base
Memory, 16-bit
Weights, 4-bit
Zuse-27B
Qwen3.8-27B
65.5 GB
15.1 GB
Zuse-9B
Qwen3.5-9B-Base
21.9 GB
5.6 GB
Zuse-4B
Qwen3.5-4B-Base
14.3 GB
3.1 GB
Table 1: The Zuse model family. Memory is resident memory when served.
Suite
Content
Size
Use during development
European suite
11 sources excluded from training, 9 task groups, 27 languages
20,044 rows
Model selection, error diagnosis, design of later runs
Transfer v4 and v9 [ 2 ]
Public English suites
764 rows in Transfer v4 per split
Development split used for selection and data choices; test split scored on five 27B checkpoints
SemIf [ 2 ]
Authored English decisions plus perturbations
144 and 108 rows
Monitored during development
Breadth report [ 2 ]
14 public datasets in 5 areas
1,990 records, development split
Scored after freezing
When2Call [ 20 ]
Tool-use choice
3,423 tasks
Scored after freezing
World languages
Belebele [ 11 ] and SIB-200 [ 12 ]
210 languages, 32,700 rows
Scored after freezing; European rows overlap the European suite’s sources
Table 2: Evaluation suites and their use during development.
Suite and split
Zuse-27B
Zuse-9B
Zuse-4B
Jev
Kev-27B
Kev-9B
Kev-4B
Transfer v4, development
0.832 / 0.246 / 0.576
0.772 / 0.316 / 0.410
0.755 / 0.342 / 0.380
0.857 / 0.211 / 0.70
0.851 / 0.225
0.820 / 0.289
0.817 / 0.269
Transfer v4, test
0.860 / 0.204 / 0.776
0.797 / 0.279 / 0.598
0.775 / 0.313 / 0.512
–
0.889 / 0.156
0.852 / 0.217
0.838 / 0.242
Transfer v9, development
0.786 / 0.305 / 0.462
0.716 / 0.382 / 0.302
0.695 / 0.410 / 0.305
–
–
–
–
Transfer v9, test
0.787 / 0.285 / 0.609
0.723 / 0.369 / 0.485
0.694 / 0.415 / 0.392
–
–
–
–
SemIf, 144 authored rows
0.924
0.847
0.882
0.965
0.903
–
–
Table 3: English public suites. Cells show accuracy / Brier / auto@5 where available; Kev cells show accuracy / Brier. Jev and Kev figures are as reported by Kev [ 2 ] .
Area
Jev
Kev-27B
Z-27B
Z-9B
Knowledge, reasoning
37.7
33.8
36.0
24.6
Language understanding
77.4
75.6
69.7
57.8
Retrieval, classification
75.4
62.9
63.0
57.7
Tools, automation
63.6
64.2
59.0
58.5
Arts, human taste
12.7
14.7
10.0
6.7
Overall
53.3
50.2
47.5
41.1
Table 4: Breadth report on the development split, by area.
Measure
Zuse-27B
Zuse-9B
Zuse-4B
Accuracy (%)
87.5
84.1
81.2
Brier score
0.185
0.229
0.264
Expected calibration error
0.016
0.023
0.032
Automatable at 1 / 2 / 5 / 10% error (%)
40 / 58 / 77 / 94
30 / 50 / 67 / 86
26 / 42 / 61 / 78
Order flips, reversed order (%)
5.6
6.2
8.9
Order flips against a third order, single / two averaged (%)
3.7 / 3.3
4.4 / 3.5
6.9 / 5.2
Table 5: European suite. Order flips use 1,223 choice rows and planted instructions a 1,500-row sample.
Figure 2: Accuracy and in-sample automation share by task group, Zuse-27B.
Figure 3: Accuracy by language on the European suite, all three models.
Group
Reading languages
Topic languages
Zuse-27B
Zuse-9B
Zuse-4B
In the list
97
139
85.1 / 87.8
76.5 / 86.1
71.0 / 84.6
Outside the list
25
66
60.5 / 65.2
48.7 / 59.6
44.8 / 56.2
Table 6: Accuracy (%) on world languages, by whether the language is in the Qwen list. Cells show reading / topic.
Model
2 questions (ms)
6 questions (ms)
5 questions, 2,200 tokens (ms)
Requests per second, 1 / 32 clients
Zuse-27B
32 / 18
45 / 31
182 / 51
20 / 47
Zuse-9B
13 / 8
17 / 12
86 / 20
49 / 118
Zuse-4B
10 / 6.5
14 / 10
84 / 18
62 / 141
Table 7: Model time per request, new document / cached document, and throughput.
Typed Decision Models such as Jev, Clef and pplx-decider return a probability for each answer an application provides. Applications use them in many languages, but are compared mostly in English. We introduce XDecideBench, a multilingual benchmark of nine suites in 122 languages. On our twelve-language suite, three open models fine-tuned for typed decisions beat their Qwen base models in every language. On new test data with local base models, their accuracy gain is larger in languages with less web text than in those with more. They fix a similar share of base-model errors in both, so the larger gain may reflect weaker base models. The best model on the twelve-language suite is best on only two other suites. The best English model on each suite, chosen in hindsight, is significantly beaten in about a quarter of suite and language combinations. On Belebele, Decision-1 is best in English but falls far behind Jev in Burmese. A confidence threshold fitted on English to allow 5% errors gives a higher error rate on XNLI, even in German and French. XDecideBench shows that results from one suite or from English are a poor guide to how a model and its threshold behave in other languages.
Marek Šuppa, Jaroslav Kopčan
Comenius University in Bratislava · Cisco Systems · Kempelen Institute of Intelligent Technologies
Many applications need either a probability distribution over a few decisions or a set of typed spans from a text, and both are usually obtained by generating text. We present Chinese Decision, a family of Chinese models that answers both with a single forward pass and no generated token: candidate-letter logits give probabilities for Boolean, categorical, and ordinal decisions, and a small span head reading an echoed copy of the input gives typed character spans. Both readouts share one backbone and one merged LoRA, so a unified model is one weight file. We release decision-only and unified models at 0.8B, 2B, and 4B (Qwen3.5), trained for one epoch on 13.05 million decisions and, for the unified models, on CJ-IE, a 1.2-million-row Chinese extraction mixture. On the full 307,900-decision CJ-Bench, the unified 4B model reaches 86.86% accuracy, above every reported baseline and the decision-only 4B model (85.97%), and extracts spans on CJ-IE with micro F1 of 68.9. Retention has a price: against a 0.8B model continued on CJ-IE alone, which reaches 73.9 F1 but forgets CJ-Bench (71.02%), keeping decision ability costs 5.0 to 10.7 F1 points, shrinking with size. Gains of the unified models over decision-only ones are small, and all comparisons rest on single training runs with paired clustered bootstrap intervals.
Qiang Yan, Jiafeng Guo
Institute of Computing Technology, Chinese Academy of Sciences
Decision models such as Jev score candidate answers or actions. A small decision head trained label-free can recover about 99% of a frozen language model's multiple-choice accuracy while using only two-thirds of its transformer layers. It learns the full model's answer probabilities with KL divergence, without correct-answer labels for training. One backbone pass reads the question and all options together; the head returns probabilities for every option. We call this model FunDM (Fun Decision Model). Using 24 of Qwen3-8B's 36 layers and skipping the LM head, its +1T variant reaches 75.0% mean accuracy over eight benchmarks, versus 75.7% for the full model. On MMLU-Pro, both heads achieve 1.56 to 1.57 times batched vLLM throughput with about one-third lower peak GPU allocator memory. Both heads score higher than released Contrastive Language Models (CLM), fine-tuned CLM, and CLM-shaped heads trained from scratch on the eight multiple-choice benchmarks and coding-agent selection. Separately trained agent heads learn judge probabilities without pass/fail labels and select passing runs for 74.4 to 74.6 of 100 held-out tasks on average, versus 74 for the full-model judge and 52.0 for a CLM-shaped action-trained head. Code is available at: https://github.com/Ujjawal-K-Panchal/FunDM