Poro 2
Finnish and English instruction-following model based on Llama 3.1
Poro 2 is a chatbot model fine-tuned for Finnish and English conversations. It was built by AMD Silo AI, TurkuNLP, and HPLT using the LUMI supercomputer. The 70B variant adds strong local language skills to the Llama 3.1 base.
- Built on Llama 3.1 70B architecture
- Trained on 165 billion tokens including Finnish
- Outperforms Llama 3.3 in Finnish benchmarks
- Released under Llama 3.3 Community License
- Developed using LUMI supercomputer resources
What it is
Poro 2 is a family of language models designed for conversational tasks in Finnish and English. The 70B Instruct version serves as a chatbot capable of following complex directions. It builds upon the Llama 3.1 architecture but undergoes extra training to master local linguistic nuances. Developers created it to support high-quality dialogue, translation, and content generation for users in Finland and beyond. The name refers to the Finnish word for reindeer, reflecting its cultural roots.
How it came about
The project emerged from a partnership between AMD Silo AI, the TurkuNLP group at the University of Turku, and High Performance Language Technologies. Researchers used the LUMI supercomputer, with resources provided by CSC IT Center for Science. The team focused on adding new language capabilities to existing open models through continued pretraining. This approach allowed them to maintain strong performance in English while significantly boosting proficiency in Finnish.
Inside the model
The base model started with Llama 3.1 70B and underwent continued pretraining on 165 billion tokens. This dataset included Finnish and English text, along with code and mathematics. Supervised fine-tuning followed, using 1.4 million instruction examples from sources like Tulu 3 and OASST2. The final step applied Direct Preference Optimization with the HelpSteer3 dataset to align responses with human preferences. The context window supports sequences up to 4,096 tokens long.
What it can do
Tests show the model outperforms Llama 3.1 and 3.3 in Finnish instruction following. It achieves a 70.79 score on IFEval Finnish and leads in MTBench Finnish evaluations. In English, it remains competitive, matching or exceeding the baseline Llama 3.3 in several metrics. Pairwise comparisons indicate a 66 per cent win rate against Llama 3.3 in Finnish conversations. The system handles multi-turn dialogue and translation tasks effectively.
How to use it
Weights are available for download under the Llama 3.3 Community License. Users can run the model locally or integrate it into applications via standard transformer libraries. The license requires reviewing terms before deployment. The developers also released the instruction collection used during training for further research. This open access supports enterprise and academic use cases requiring strong multilingual support.
Safety and testing
The creators note that the model may reflect biases present in its training data. Outputs can sometimes be inaccurate, prejudiced, or controversial. Proficiency drops significantly for languages other than English and Finnish. Users should expect potential errors in factual information. The team highlights these limitations to encourage responsible deployment in customer service and creative writing contexts.
Why it matters
Poro 2 demonstrates how continued pretraining can efficiently adapt large models to specific languages. It provides a robust baseline for Finnish natural language processing tasks. By releasing both base and instruct versions, the team supports further customization by the community. The project highlights the value of regional supercomputing resources in developing specialized AI tools.
Questions people ask about Poro 2
What languages does Poro 2 support?
The model is optimized for Finnish and English conversations. It has limited proficiency in other languages. Developers trained it specifically to handle translation and dialogue in these two primary tongues.
Who developed Poro 2?
AMD Silo AI, the TurkuNLP group from the University of Turku, and High Performance Language Technologies collaborated on the project. They used computing resources from CSC IT Center for Science.
Is Poro 2 free to use?
Yes, the weights are open for download. It uses the Llama 3.3 Community License. Users must agree to the license terms before using the model for commercial or research purposes.
Good reading and sources
There is no English Wikipedia article on Poro 2 yet, so the facts on this page come from the sources listed above: huggingface.co/LumiOpen/Llama-Poro-2-70B-Instruct. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.