Meet Zuse, the decision model.Like Jev, but built and hosted in Europe.

Zuse reads a document, picks from the options you set and tells you how sure it is.

Get an API key

Live· real answers from Zuse 27B

Customer email137 / 4,000
  1. ✓ Act
    • Billing96.5%
    • Technical0.0%
    • Travel0.2%
    • Other3.3%
  2. ✓ Act
    • Yes95.8%
    • No4.2%
  3. ⚑ Check
    • Not urgent0.5%
    • Normal3.4%
    • Urgent89.5%
    • Critical6.6%

✓ Act at 90% sure or more, ⚑ Check below it. You choose your own level.

Saved answer

Zuse in numbers

Tested on tasks it never trained on, and against the chat models you already know.

$0.12

Zuse 9B against $0.34 for GPT-5.6-Luna, 2.9× cheaper

What 1,000 decisions on a one-page document cost. Chat models are counted at 1,200 input and 60 output tokens a decision, with no reasoning. Lower is better.

Zuse 9Bbest$0.12
GPT-5.6-Luna$0.34
Zuse 27B$0.75
Claude Haiku 4.5$1.65
Claude Sonnet 5$3.30
Claude Opus 5$8.25
List prices, read live from the Bineric API.
65,536
tokens per document on the API, about 100 pages of text
0.2 s
typical time to answer through the API, every question at once
1.6%
gap between how sure it says it is and how often it is right

How it works

  1. 1Send the documentAn email, a form, a clause, a case note, in 27 European languages and many more.
  2. 2Ask your questionsWrite any question and list the answers it may pick from: yes or no, a scale, or up to 255 options.
  3. 3Act on the answerEach answer comes with how sure Zuse is, so you choose which ones to act on automatically.

The document

“Hi, I was charged twice for my monthly pass last week. I want my money back as soon as possible, this is the second time it has happened!”
Which team should handle it?
Billing 99.1%
Is the customer asking for a refund?
Yes 98.7%
How urgent is it?
Urgent 48.6%Critical 41.2%
Example values. Three questions, one pass, 214 input tokens.

What can you use Zuse for?

Anywhere a person reads something and picks from a fixed set of answers.

Customer service

Send each email, chat or form to the right team, and flag the urgent ones.

Which team should handle this?

  • Billing
  • Technical
  • Sales

Claims and casework

Sort incoming claims, applications and cases, and check they are complete.

Is a receipt attached?

  • Yes
  • No
  • Unclear

Contract and policy checks

Find out whether a clause is there and which terms apply.

Does the contract state a notice period?

  • Yes
  • No

Reviews and surveys

Read thousands of free-text answers and count what people talk about.

What is this review about?

  • Delivery
  • Product
  • Price
  • Service

Content moderation

Check posts and comments against your rules, and flag the unsure ones.

Does this post break the rules?

  • No
  • Review
  • Remove

Guardrails for agents

Check an agent's plan against a policy before it acts.

Does this refund follow the policy?

  • Yes
  • No

Where Zuse does best

  • 94%

    What a customer wants

  • 93%

    Questions about a document

  • 90%

    Happy or angry

  • 90%

    Sorting by topic

Share of 20,044 questions answered correctly, in 27 languages, on tasks Zuse never trained on.

Pricing

Priced per input token. Zuse writes no text, so there is nothing else to pay for.

Zuse 27B

zuse-27b

Highest accuracy, for the decisions you automate.

$0.75 / 1M input tokens

About $0.75 per 1,000 one-page decisions

  • No output tokens to pay for
  • 27 European languages and many more, hosted in the EU
  • Same key as the Bineric API
Get an API key

Zuse 9B

zuse-9b

Fast and low cost, for routing at volume.

$0.12 / 1M input tokens

About $0.12 per 1,000 one-page decisions

  • No output tokens to pay for
  • 27 European languages and many more, hosted in the EU
  • Same key as the Bineric API
Get an API key

Dedicated or on-premises

For public sector and regulated companies.

Custom

Priced for your volume

  • A GPU reserved for you in an EU data centre
  • Your own endpoint, no shared traffic
  • No per-token billing
  • Data processing agreement included
  • Or run it on your own servers
  • Or in your own AWS account, from AWS Marketplace
Talk to us

Questions

What is Zuse?

A model that reads a document and answers your questions about it with a short answer, picked from options you choose, and tells you how sure it is. Use it to sort emails, check forms, flag urgent cases and more.

How is Zuse different from a chat model?

A chat model writes a long answer that you then have to read and check, and you pay for every word. Zuse picks from the answers you give it and says how sure it is. It never writes text of its own, so it is faster and cheaper, and it cannot wander off topic.

How does Zuse compare with Jev?

It answers the same kind of questions as TypeSafe's Jev, with the same request and response format. On Jev's own public English benchmark, Zuse 27B matches its accuracy (86.0% against 85.7% published) and can safely automate more decisions: 77.6% against 70.0%, with half as many confident mistakes. And it runs in the EU, is trained on 27 European languages and comes with a data processing agreement.

Zuse and Jev side by side, and how to move your Jev code →
Which languages does it support?

115 of the 210 languages we tested score 85% or more, 49 of them European: Afrikaans, Albanian, Arabic, Armenian, Assamese, Asturian, Awadhi, Azerbaijani, Banjar, Basque, Belarusian, Bengali, Bhojpuri, Bosnian, Bulgarian, Cantonese, Catalan, Cebuano, Chhattisgarhi, Chinese (Simplified), Chinese (Traditional), Crimean Tatar, Croatian, Czech, Danish, Dari, Dutch, Egyptian Arabic, English, Esperanto, Estonian, Faroese, Finnish, French, Galician, Georgian, German, Greek, Hebrew, Hindi, Hungarian, Icelandic, Indonesian, Irish, Italian, Japanese, Javanese, Kannada, Kashmiri (Arabic script), Kazakh, Khmer, Korean, Kyrgyz, Latvian, Levantine Arabic, Ligurian, Limburgish, Lithuanian, Lombard, Luxembourgish, Macedonian, Magahi, Malay, Malayalam, Maltese, Marathi, Mesopotamian Arabic, Minangkabau, Najdi Arabic, Nepali, Norwegian Bokmål, Norwegian Nynorsk, Occitan, Odia, Papiamento, Persian, Polish, Portuguese, Punjabi, Romanian, Russian, Samoan, Sanskrit, Sardinian, Scottish Gaelic, Serbian, Sicilian, Sinhala, Slovak, Slovenian, South Azerbaijani, South Levantine Arabic, Spanish, Sundanese, Swahili, Swedish, Ta'izzi-Adeni Arabic, Tagalog, Tajik, Tamil, Tatar, Telugu, Thai, Tok Pisin, Tunisian Arabic, Turkish, Turkmen, Ukrainian, Urdu, Uyghur, Uzbek, Venetian, Vietnamese, Welsh, Yiddish. Zuse was trained on 27 European languages; the rest comes from its base model. Just under the bar: Friulian (83%), Silesian (83%), Latgalian (80%).

How accurate is it?

On 20,044 questions in 27 languages, from tasks it never trained on, Zuse 27B answers 87.5% correctly. On routing, reading and sentiment it is right 90 to 95% of the time. Every answer comes with how sure Zuse is, so you can act automatically on the ones it is sure about.

Which model should we choose?

Zuse 27B is the most accurate, for the decisions you automate. Zuse 9B is faster and cheaper, for sorting at high volume. Both use the same API and the same key, so you can start with one and switch.

What does it cost?

You pay only for the text Zuse reads. It writes no text, so there are no output tokens to pay for: a decision on a one-page document costs a fraction of a cent. Prices for each model are in the pricing section above.

Where is it hosted, and what happens to our data?

Zuse runs on servers in the EU (Spain, Germany or Sweden, and another EU region when GPU capacity requires it). Your documents and questions are not stored, not logged and never used for training; we keep only metadata: token counts and cost for billing, and counts such as the number of questions and the document's language. Our data processing agreement and list of subprocessors are at bineric.com/legal.

Can we run it on our own servers?

Yes. A dedicated instance reserves a GPU in an EU data centre for you alone, with its own endpoint, and we can also run Zuse on your own servers. Talk to us and we will price it for your volume.

When should we check its answers?

Set a confidence level per question and look at anything below it. Routing, reading and sentiment automate well; legal, grammar and general-knowledge questions are better checked. Do not use Zuse alone for decisions about a person's employment, credit, benefits or legal rights.

Easy to get started with Zuse

Get an API key, send a document and your questions, and the answers come straight back.

Your first request
curl https://api.bineric.com/v1/zuse \
-H "Authorization: Bearer $BINERIC_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "zuse-27b",
"state": "Hi, I was charged twice for my monthly pass ...",
"questions": {
"refund": {
"type": "noul",
"instructions": "Is the customer asking for a refund?"
}
}
}'