For those curious, the 24 official languages are Bulgarian, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Greek, Hungarian, Irish, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Slovak, Slovenian, Spanish, and Swedish.
Maltese, interestingly, is the only Afro-Asiatic derived language.
Hungarian, Finnish, and Estonian are the three Uralic languages.
All the others are Indo-European, Greek being the only Hellenic one, Irish the only Celtic, the rest are Baltic, Slavic, Italic, or Germanic.
(I originally used the term Balto-Slavic, though I was unaware of some of the connotations of that term until just now. Baltic and Slavic do share a common origin, but that was a very very long time ago)
I was thinking the same, why are so many superior models coming from only countries like US and China. And why are European countries not in the list other than France with Mistral. Why are so few companies in India, Japan, South Korea even close to a promising new model like what Chinese companies did ?
Aren't all frontier models already able to use all these languages? Support for specific languages doesn't need to be built in, LLMs support all languages because they are trained on multilingual data.
Term support is vague. Can you do basic interaction in most other languages? Sure. Is it anywhere close to competence it has in english? No. Most models seem to just translate english responses at beginners simplistic monotone level.
>The EuroLLM Team brings together some of the brightest minds in AI including Unbabel, Instituto Tecnico Lisbon, the University of Edinburgh, Instituto de Telecommunicacoes, Université Paris-Saclay, Aveni, Sorbonne University, Naver Labs, and the University of Amsterdam.
>Europe is the only continent in the world to have a large public network of supercomputers that are managed by the EuroHPC Joint Undertaking (EuroHPC JU). As soon as we received the EuroHPC JU access to the supercomputer, we were ready to roll up our sleeves and get to work. We developed the small model right away and in less than 6 months the second model was ready.
>Europe is the only continent in the world to have a large public network of supercomputers that are managed by the EuroHPC Joint Undertaking (EuroHPC JU).
Who would have thought that Europe is the only continent to have a network of supercomputers managed by Europe⸮
If I want to use an LLM to do translation, should I use a base model or an instruction tuned version? I've had mixed results using the chat models and a simple "Translate this to <language>: "
As expected, Europe finally catches up to 2024 and launches an LLM that barely competes against the heavyweights.
The US and China are running rings around Europe.
Mistral is an exception as it was funded by US VCs and they are a great example showing that without VC funding, Mistral would have been begging to the EU for a microsopic grant to train a LLM worse than Llama.
I'm somewhat skeptical of taxpayer funded innovation. Seen a few Horizon grants from the side, as a citizen I'd prefer to not pay for them, but unfortunately can't opt out.
It seems like it, in most ways, it would be bad to train on 24 separate languages. That's just 24 partitions to the data. Seems really inefficient and better to simply train in the biggest (english) and translate.
I do think this will introduce some biases that correlate with the English language. It would be interesting to see more specifically what this means. But regardless, I don't think you can produce a competitive model with such a large subdivision of training data.
It’s about control. It is a fundamental difference in perspective and mentality regarding control. America is or at least was largely oriented around indirect control or outcome oriented objectives, where Europe is largely oriented around direct control or prescribed objectives.
It’s also enshrined in our respective governments and the foundational philosophies that underpin them. The US Declaration of Independence sets out to describe that the natural rights of the men who created the USA are preeminent and the Constitution lists how those rights may not be infringed upon, i.e. It creates laws that binds and limits the actions of government, something that was and has never been emulated since. Where across Europe, you simply do not have anything even remotely similar and the law inversely describes what you are permitted by government to do instead.
It is effectively descriptive vs prescriptive law and underlying philosophies. It is something I have had the hardest time on occasion getting my European friends to really internalize, seemingly because it’s so contrary to what their conditioned with all their life, i.e., the government is essentially the matter that grants you what it grants you, not that you have rights that the government may not infringe upon.
But to be fair, this possibly European tendency to dominate and control what you may and may not do and when and how, is and long has been creeping into the USA too, arguably since even the 12th amendment to the Constitution and getting worse with every amendment since, layers upon layers of contradicting and conflicting flaws and bugs that will be reading their ugly heed here in about two years, when Trump may run for President again. And if you don’t think he can, you simply don’t understand what a spaghetti code the Constitution is after the 11th amendment, hack after hack building up mountains of debt that is going to come due in our lives.
Please don't fulminate on HN. A sentence like "Actually nuts to me the degree to which European policymakers do not even begin to understand" is inflammatory rhetoric of the kind we're trying to avoid here on HN. The question of how countries/economic unions can funtion most effectively is a topic worthy of serious discussion, but these discussions can be far more fruitful it they're approached with curiosity rather than rage.
I used the 9B Instruct version, from the small models, it was the one with the best Latvian knowledge out there, bar none. GPT-OSS 20B and Qwen3 30B A3B and similar ones weren't even close.
That said, the model itself was a little bit dumb and not something you'd really use for programming/autocomplete or tool calling or anything like that, which also presented some problems - even for processing text, if you need RAG or tool server calls, you need to use something like Qwen3 for the actual logic and then pass the contents to EuroLLM for translation/formatting with the instructions, at which point your n8n workflow looks a bit messy and also you have to run those two models instead of only one.
Meanwhile, the best cloud model for Latvian that I've found so far was Google Gemini 2.5 Pro, but obviously can't use cloud models in certain on-prem use cases.
78 comments
[ 1.3 ms ] story [ 86.8 ms ] threadMaltese, interestingly, is the only Afro-Asiatic derived language.
Hungarian, Finnish, and Estonian are the three Uralic languages.
All the others are Indo-European, Greek being the only Hellenic one, Irish the only Celtic, the rest are Baltic, Slavic, Italic, or Germanic.
(I originally used the term Balto-Slavic, though I was unaware of some of the connotations of that term until just now. Baltic and Slavic do share a common origin, but that was a very very long time ago)
Edit: Thanks, @Bengalilol.
The 1.7B one looks meh.
But really solid numbers on the 9B! Props to the team!
"all official 24 EU languages” to "all 24 official EU languages"
>Europe is the only continent in the world to have a large public network of supercomputers that are managed by the EuroHPC Joint Undertaking (EuroHPC JU). As soon as we received the EuroHPC JU access to the supercomputer, we were ready to roll up our sleeves and get to work. We developed the small model right away and in less than 6 months the second model was ready.
[1] https://www.eurohpc-ju.europa.eu/eurohpc-success-story-speak...
Repurposing some of that physics sim compute
Who would have thought that Europe is the only continent to have a network of supercomputers managed by Europe⸮
0: https://sites.google.com/view/eurollm/home
The US and China are running rings around Europe.
Mistral is an exception as it was funded by US VCs and they are a great example showing that without VC funding, Mistral would have been begging to the EU for a microsopic grant to train a LLM worse than Llama.
It seems like it, in most ways, it would be bad to train on 24 separate languages. That's just 24 partitions to the data. Seems really inefficient and better to simply train in the biggest (english) and translate.
I do think this will introduce some biases that correlate with the English language. It would be interesting to see more specifically what this means. But regardless, I don't think you can produce a competitive model with such a large subdivision of training data.
2. A credible scale effort for EU own silicon for AI Compute, wouldn't hurt either.
3. And this can only be achieved by vertical integration to combat fragmentation.
It’s also enshrined in our respective governments and the foundational philosophies that underpin them. The US Declaration of Independence sets out to describe that the natural rights of the men who created the USA are preeminent and the Constitution lists how those rights may not be infringed upon, i.e. It creates laws that binds and limits the actions of government, something that was and has never been emulated since. Where across Europe, you simply do not have anything even remotely similar and the law inversely describes what you are permitted by government to do instead.
It is effectively descriptive vs prescriptive law and underlying philosophies. It is something I have had the hardest time on occasion getting my European friends to really internalize, seemingly because it’s so contrary to what their conditioned with all their life, i.e., the government is essentially the matter that grants you what it grants you, not that you have rights that the government may not infringe upon.
But to be fair, this possibly European tendency to dominate and control what you may and may not do and when and how, is and long has been creeping into the USA too, arguably since even the 12th amendment to the Constitution and getting worse with every amendment since, layers upon layers of contradicting and conflicting flaws and bugs that will be reading their ugly heed here in about two years, when Trump may run for President again. And if you don’t think he can, you simply don’t understand what a spaghetti code the Constitution is after the 11th amendment, hack after hack building up mountains of debt that is going to come due in our lives.
Ah the good old "Europe can't do Silicon Valley" trope.
https://news.ycombinator.com/newsguidelines.html
I used the 9B Instruct version, from the small models, it was the one with the best Latvian knowledge out there, bar none. GPT-OSS 20B and Qwen3 30B A3B and similar ones weren't even close.
That said, the model itself was a little bit dumb and not something you'd really use for programming/autocomplete or tool calling or anything like that, which also presented some problems - even for processing text, if you need RAG or tool server calls, you need to use something like Qwen3 for the actual logic and then pass the contents to EuroLLM for translation/formatting with the instructions, at which point your n8n workflow looks a bit messy and also you have to run those two models instead of only one.
Meanwhile, the best cloud model for Latvian that I've found so far was Google Gemini 2.5 Pro, but obviously can't use cloud models in certain on-prem use cases.