11h ago

Europe will have its own AI: OpenEuroLLM

www.heise.de Language model: OpenEuroLLM aims to make AI in EU more independent and diverse

A European consortium wants to build an open source family of powerful, multilingual large language models for private and public services.

The European Comission has granted €20 million to STEP, the European consortium that will create the AI model. It will be open source, European regulations-compliant and unlike Deepseek, its dataset will also be open source and will be trained in 35 languages.

27 comments

with that little money split between that many institutions, nothing will come of it.
It's especially pointless ever since DeepSeek R1 dropped. Now everyone has the recipie to build state of the art models, so it's only a matter of time until European companies will create one.

I am conflicted about this choice. I am happy that the EU Commission will invest funds into open source technologies, but at the same time the US and China are already investing enough into "free as in free beer" models. Is it really worth it building yet another model?
Why not fund open source software development instead of funding machine learning? €20 million would do miracles divided between a few teams of developers, but they might merely be bread crumbs for machine learning training.
- 2 main issues with the lack of Euro models: 1) Performance of all SOTA models is much better in English. 2) US models have US values. It's yet another tool to culturally assimilate Europe (and the rest of the world too)
- Is it really worth it building yet another model?
  Yes, it is, and it has to do with independence and many other reasons. It'll be multilingual, legally compliant, it comes without Chinese nor other censorship, it is open source unlike Deepseek, ChatGPT, and others.
  
  Mmh, okay that makes sense. Especially the multilinguality would be pretty important. As for the legality, we'll see how it goes. Do we even know if it's really possible to build a good model with only legally acquired data?
  As for the censorship, as far as I know, for DeepSeek's models it's injected in the prompt after the training is completed, so it shouldn't really be censored if you run it locally.
  But yeah, you have raised good points. Thanks.
- That money should definitely go towards funding sovereign cloud infrastructure and open source software instead of vaporware AI bullshit. Where will you run your LLMs if you have no infra..
  
  We already have sovereign Cloud Infrastructure (OVH, Scaleway etc...)
  Most people use AWS, Azure and Google Cloud because of Resume Driven Development and nobody got fired for buying AWS, and most of them probably don't need them

Hey! Look at me! I also want to be a dumpster fire.
- At the very least the name is much more technically accurate by putting LLM instead of AI in there.

Well, investing a mere €20 millions won't achieve much. On the other hand I'm glad they aren't wasting more money on it.
- Wasn't DeepSeek v3 trained with single-digit million dollars budget?
  
  Iirc leaked numbers says something closer to 1 billion USD
  
  Probably not. There's a lot of reasons to be skeptical of those claimed numbers.

Who actually wants AI?
- To be fair — I wouldn’t mind an open source AI model that works decently well and isn’t made and selectively propagandised by China or Meta.
  Obviously I’m sick of all the LLM enshittification — but there’s a couple tasks I wouldn’t mind having a FOSS LLM for.
  
  I could get down with that.
- I mean a whole lot of people are using AI for a whole lot of different things. It's easy to hate on, but it's here to stay.
- I do!
  
  Me too

Have all of these people forgotten what Open is supposed to mean?

27 comments