Finnish AI Models: A Practical Guide

Finnish is one of the hardest languages for general-purpose AI. Long compound words, 15 grammatical cases and heavy inflection mean that models trained mostly on English often stumble, producing stiff phrasing or outright grammar errors. A handful of Finnish and Nordic research groups have spent the last few years fixing that, and the result is a solid set of open models you can actually download and use. Here is what exists, what each one does well, and where it fits.

The Models at a Glance

Model	Developer	Type	Languages	Best For
Poro 34B	Silo AI and TurkuNLP (University of Turku)	Large language model	Finnish, English, code	Finnish text generation and chat
Viking (7B, 13B, 33B)	Silo AI and TurkuNLP	Large language model	Finnish, English, Swedish, Danish, Norwegian, Icelandic, code	Nordic multilingual work
Poro 2	Silo AI and TurkuNLP	Instruction-tuned LLM	Finnish, English	Assistants and applied tasks
FinBERT	TurkuNLP	Encoder (BERT)	Finnish	Classification and entity recognition
FinGPT	TurkuNLP	Generative model family	Finnish	Research and small-scale generation
OPUS-MT (fi-en and others)	Helsinki-NLP (University of Helsinki)	Machine translation	Finnish and many others	Translation pipelines
Finnish speech recognition	Aalto University speech research	Speech-to-text	Finnish	Transcription and voice tools

Poro 34B

Poro 34B was one of the first large models built with Finnish as a core language rather than an afterthought. It has 34 billion parameters and was trained on roughly a trillion tokens of Finnish, English and programming code, using the LUMI supercomputer. The mix of Finnish and English is deliberate: the English data helps the model reason and follow structure, while the Finnish data teaches it natural grammar and vocabulary.

Features

• 34 billion parameters, trained on Finnish, English and code
• ALiBi positional encoding, which helps with longer inputs
• Released openly under a permissive Apache 2.0 licence
• Includes a bilingual translation capability between Finnish and English

Strengths

• Fluent, grammatically careful Finnish output
• Strong base for fine-tuning on company or domain data
• Fully open weights, so it can run on private infrastructure

Viking

Viking widens the lens to the whole Nordic region. It comes in three sizes (7B, 13B and 33B parameters) and covers Finnish, Swedish, Danish, Norwegian, Icelandic, English and code. Because these languages share vocabulary and structure, the model benefits from cross-lingual transfer: what it learns in Swedish helps with Danish, and so on.

Features

• Three sizes, so you can match the model to your hardware
• Six Nordic-region languages plus English and code
• Open weights under Apache 2.0

Strengths

• The smaller sizes run on modest GPUs
• Excellent for companies serving several Nordic markets at once
• Consistent quality across closely related languages

Poro 2

Poro 2 takes a different route. Instead of training from scratch, it builds on Meta’s Llama 3.1 models and adds extensive Finnish training and instruction tuning on top. The result is a model that keeps the broad capability of a modern frontier-class base while speaking Finnish far better than the original.

Features

• Available in smaller and larger sizes derived from Llama 3.1
• Instruction-tuned, so it follows prompts and holds a conversation
• Bilingual Finnish and English performance

Strengths

• Much better at following instructions than the earlier base models
• Practical for chatbots, summarisation and customer-facing assistants
• Benefits from the strong reasoning of its base model

FinBERT

Not every task needs a giant generative model. FinBERT, from the Turku NLP group, is a BERT-style encoder trained on billions of words of Finnish text from news, online discussion and other sources. It does not write text; it understands it.

Features

• Cased and uncased versions
• A custom vocabulary built for Finnish word forms
• Widely available on Hugging Face

Strengths

• Small, fast and cheap to run
• Very good at sentiment analysis, topic classification and named entity recognition
• Easy to fine-tune with limited data

FinGPT

FinGPT is a family of generative Finnish models, also from TurkuNLP, ranging from small models of a couple of hundred million parameters up to billions. They were created to give researchers a transparent, well-documented set of Finnish-first generative models.

Features

• Multiple sizes in a single family
• Trained on a large, carefully documented Finnish corpus
• Open for research use

Strengths

• Useful for experiments where you need to understand exactly what data the model saw
• Lightweight options for devices with limited memory

OPUS-MT

The University of Helsinki’s language technology group built OPUS-MT, a large collection of translation models trained on open parallel data. The Finnish-English pairs are among the most used, and the models are compact enough to run on a regular laptop.

Features

• Hundreds of language pairs, with Finnish in several directions
• Built on the Marian neural translation framework
• Free to download and self-host

Strengths

• Fast and inexpensive for bulk translation
• Runs offline, which matters for confidential documents
• Easy to integrate into existing content pipelines

Finnish Speech Recognition

Aalto University’s speech group has long been a leader in Finnish speech-to-text. Finnish is written almost exactly as it is spoken, but its endless word forms make vocabulary-based recognisers struggle. Aalto’s models handle this with subword modelling, and the work has been strengthened by large crowdsourced speech collections.

Features

• Subword-based recognition suited to long compound words
• Trained on large volumes of spontaneous Finnish speech

Strengths

• Accurate transcription of real-world, conversational Finnish
• Good foundation for voice assistants, subtitling and call analysis

Picking the Right One

For general writing and chat in Finnish, start with Poro 2 or Poro 34B. If your business spans several Nordic countries, Viking is the more sensible choice. For classification, search and sentiment work, FinBERT will do the job faster and at a fraction of the cost. Translation workloads belong with OPUS-MT, and anything involving audio points to the Aalto speech models. Since most of these are open and Apache-licensed or similarly permissive, the best approach is often to test two or three on your own data and keep whichever performs best.