AI for Hungarian law: introducing KTwo Mantle 0.5

General-purpose AI models are trained mostly on English. Ask one about Hungarian law and it has to work in a language it saw relatively little of, in a legal system it saw even less of, about rules that may have changed since its training data was collected. KTwo Mantle 0.5 is KTwo’s model for this domain: a 30B model trained on Hungarian law that, like Crust 1 Beta, trains itself. This post explains why the problem needs a dedicated model, how to judge any legal model, and where Mantle stands today.

Why general models struggle with Hungarian law

Hungarian is a small language online

Hungarian has around 13 million speakers, a small share of the text on the web and in the datasets models are trained on. Hungarian legal text, with its own vocabulary and conventions, is a small share of that. A general model has simply seen far less of it than of English legal writing.

The language is built differently

Hungarian is agglutinative: grammar is carried by suffixes stacked onto a stem, so a single word can say what English needs a phrase for. Jogszabályváltozásokról means “about changes in legislation”. Tokenizers built mostly on English break words like that into many fragments, which makes the same text longer, costlier to process and harder for the model to handle. In our measurement, the same sentence took about twice as many tokens in Hungarian as in English; see What is a token?

Legal Hungarian has its own conventions

Hungarian legislation is cited by year, number and type. The Civil Code is Act V of 2013 on the Civil Code (a Polgári Törvénykönyvről szóló 2013. évi V. törvény), usually shortened to Ptk., and its sections are numbered by book and section: the general rule of liability for damage caused outside a contract is Ptk. 6:519. §. A model has to know these conventions to read a question correctly and to cite its answer in a form a Hungarian lawyer would accept.

The law keeps changing

Hungary’s main codes are recent, and they keep moving. The Fundamental Law (Alaptörvény) has been in force since 1 January 2012; the current Civil Code since 15 March 2014, replacing the Civil Code of 1959; the current Criminal Code (Btk.) since 1 July 2013. New laws and amendments are published in the official gazette, Magyar Közlöny, throughout the year, and the Kúria, Hungary’s supreme court, issues uniformity decisions (jogegységi határozatok) that bind the courts. On top of that sits EU law, which has applied since Hungary joined the European Union in 2004.

A model frozen at one moment will cite repealed provisions with confidence, which is why it matters how a legal model learns. Mantle 0.5 trains itself, like Crust 1 Beta; KTwo hasn’t published what it trains on beyond Hungarian law, or how. We explain the general idea in What is a self-training AI model?

What a legal AI model is for, and what it isn’t

Used well, a legal model is a fast, tireless first reader:

• Finding the relevant rules: which act and section governs a question, and what related provisions say.
• Explaining a provision or a decision in plain Hungarian, or in English.
• First drafts of routine documents, letters and summaries, for a lawyer to review.
• Summarizing long judgments, contracts and legislative changes.

It is not a lawyer. A language model can state wrong things fluently, miss a recent amendment or misread the facts. Its output is a starting point that someone qualified has to check against the official sources, and nothing it writes is legal advice.

How to judge a Hungarian legal model

Whether you are testing Mantle 0.5 or anything else, these questions separate a useful legal model from a merely fluent one:

1. Does it cite precisely? Act, year, number and section in Hungarian form, such as Ptk. 6:519. §, and do the cited provisions exist?
2. Does it know what is in force? Ask about a rule that changed: does it answer from the current text or the old one, and does it know the difference between the 1959 and 2013 Civil Codes?
3. Does it weigh sources correctly? Legislation, Kúria uniformity decisions and individual judgments don’t carry the same weight.
4. Does it handle legal Hungarian? Terminology, long compound words and the formal register, not just everyday Hungarian.
5. Does it say when it isn’t sure? A good legal model flags uncertainty and points to what should be checked.
6. Does it keep up? Test it on changes published after a known date.

Where KTwo Mantle 0.5 stands

Mantle 0.5 has 30 billion parameters, more than three times the size of Crust 1 Beta. It is trained on Hungarian law, and it trains itself, like Crust 1 Beta. It is not publicly available yet. KTwo hasn’t published its training data beyond its being Hungarian law, and its API ID, modality and pricing haven’t been announced. Its current spec is on the Mantle 0.5 page, and the two models are compared side by side on All models.

If you work with Hungarian law and want to hear when that changes, write to us. This post is also available in Hungarian.

Frequently asked questions

Is there an AI model trained specifically on Hungarian law?

Yes: KTwo Mantle 0.5 is a 30B model trained on Hungarian law. It is not publicly available yet.

Can an AI model give legal advice?

No. A language model can help find, explain and draft, but its output has to be checked by a qualified lawyer against the official sources, and it is not legal advice.

Why do general AI models make mistakes on Hungarian law?

They see relatively little Hungarian, and even less Hungarian legal text, in training; Hungarian’s structure is hard for English-centric tokenizers; and a frozen model doesn’t know about amendments made after its training data was collected.

When will KTwo Mantle 0.5 be available?

KTwo hasn’t announced a date, and Mantle 0.5’s API ID and pricing aren’t announced yet either. Contact us to hear when that changes.