Finnish artificial intelligence startup Silo AI published a new open source large language model (LLM) on 12 November. The new model Poro, named after the typical Nordic tundra roaming reindeer, is the second significant open source LLM hailing from Europe after French Mistral AI. For the moment it covers Finnish, English and some coding languages, but Silo AI’s plan is to expand it to work for all 24 official European Union languages.
“I personally believe that eventually there’s going to be a lot of open source alternatives out there. The most secure way forward is to actually go open source and have full visibility into how these models have been built and what the architecture is”, said Peter Sarlin, Silo AI’s CEO in VentureBeat’s interview. Poro is released under the Apache 2.0 License.
According to the company, Helsinki-based Silo AI is the largest private AI lab with over 300 employees. Although it is a private company, its innovation relies on public sector cooperation and EU support. The Poro model is built by Silo AI’s generative AI arm SiloGen in collaboration with the University of Turku and the Horizon Europe-funded High Performance Language Technologies (HPLT) project. HPLT aims to combine large quantities of data from many languages. It has a total of 13 petabytes of web-crawled data and its data set for Finnish contains over 10 billion words.
Poro training is powered by Europe’s fastest supercomputer LUMI, situated in Kajaani, Finland. LUMI is funded through the EuroHPC joint undertaking and has received some additional funding from the European Regional Development Fund (ERDF).
Poro gains its competitive edge over LLMs developed for the widely-used languages via an innovative cross-training technique. The model is first fed two languages, and then it figures out the relationship between them. This allows Poro to seek answers in English even if the user has entered a prompt in Finnish.
Although Silo AI’s ambition is to make Poro a real European contender for the global big tech, it doesn’t only compete against the behemoths from the other side of the Atlantic. The model spurs French Mistral and German Aphec Alpha, and 2024 will undoubtedly introduce new open source AI competitors from other EU countries as well.
Featured photo by Philip Swinburn