\n<\/p>\n
Large language models (LLMs) landed on Europe\u2019s digital sovereignty agenda with a bang last week, as news emerged<\/a> of a new program to develop a series of \u201ctruly\u201d open source LLMs covering all European Union languages.<\/p>\n This includes the current 24 official EU languages, as well as languages for countries currently negotiating for entry to the EU market, such as Albania<\/a>. Future-proofing is the name of the game.<\/p>\n OpenEuroLLM<\/a> is a collaboration between some 20 organizations, co-led by Jan Ha<\/a>ji\u010d<\/a>, a computational linguist from the Charles University in Prague, and Peter Sarlin<\/a>, CEO and co-founder of Finnish AI lab Silo AI, which AMD acquired last year for $665 million<\/a>.<\/p>\n The project fits a broader narrative that has seen Europe push digital sovereignty as a priority, enabling it to bring mission-critical infrastructure and tools closer to home. Most of the cloud giants are investing<\/a> in local infrastructure<\/a> to ensure EU data stays local, while AI darling OpenAI recently unveiled<\/a> a new offering that allows customers to process and store data in Europe.<\/p>\n Elsewhere, the EU recently signed an $11 billion deal<\/a> to create a sovereign satellite constellation to rival Elon Musk\u2019s Starlink.<\/p>\n So OpenEuroLLM is certainly on-brand.<\/p>\n However, the stated budget<\/a> just for building the models themselves is \u20ac37.4 million, with roughly \u20ac20 million coming from the EU\u2019s Digital Europe Programme<\/a> \u2014 a drop in the ocean compared to what the giants of the corporate AI world are investing. The actual budget is more when you factor in funding allocated for tangential and related work, and arguably the biggest expense is compute. The OpenEuroLLM project\u2019s partners include EuroHPC<\/a> supercomputer centers in Spain, Italy, Finland, and the Netherlands \u2014 and the broader EuroHPC project has a budget of around \u20ac7 billion.<\/p>\n But the sheer number of disparate participating parties, spanning academia, research, and corporations, have led many to question whether<\/a> its goals are achievable. Anastasia Stasenko, co-founder of LLM company Pleias<\/a>, questioned whether<\/a> a \u201csprawling consortia of 20+ organizations\u201d could have the same measured focus of a homegrown private AI firm.<\/p>\n \u201cEurope\u2019s recent successes in AI shine through small focused teams like Mistral AI<\/a> and LightOn<\/a> \u2014 companies that truly own what they\u2019re building,\u201d Stasenko wrote. \u201cThey carry immediate responsibility for their choices, whether in finances, market positioning, or reputation.\u201d<\/p>\n The OpenEuroLLM project is either starting from scratch or it has a head start \u2014 depending on how you look at it.<\/p>\n Since 2022, Haji\u010d has also been coordinating the High Performance Language Technologies (HPLT<\/a>) project, which has set out to develop free and reusable datasets, models, and workflows using high-performance computing (HPC). That project is scheduled to end in late 2025, but it can be viewed as a sort of \u201cpredecessor\u201d to OpenEuroLLM, according to Haji\u010d, given that most of the partners on HPLT (aside from the U.K. partners) are participating here, too.<\/p>\n \u201cThis [OpenEuroLLM] is really just a broader participation, but more focused on generative LLMs,\u201d Haji\u010d said. \u201cSo it\u2019s not starting from zero in terms of data, expertise, tools, and compute experience. We have assembled people who know what they\u2019re doing \u2014 we should be able to get up to speed quickly.\u201d<\/p>\n Haji\u010d said that he expects the first version(s) to be released by mid-2026, with the final iteration(s) arriving by the project\u2019s conclusion in 2028. But those goals might still seem lofty when you consider that there isn\u2019t much to poke at yet beyond a bare-bones GitHub profile<\/a>.<\/p>\n \u201cIn that respect, we are starting from scratch \u2014 the project started on Saturday [February 1],\u201d Haji\u010d said. \u201cBut we have been preparing the project for a year [the tender<\/a> process opened in February 2024].\u201d<\/p>\n From academia and research, organizations spanning Czechia, the Netherlands, Germany, Sweden, Finland, and Norway are part of the OpenEuroLLM cohort, in addition to the EuroHPC centers. From the corporate world, Finland\u2019s AMD-owned AI lab Silo AI is on board, as are Aleph Alpha (Germany), Ellamind (Germany), Prompsit Language Engineering (Spain), and LightOn (France).<\/p>\n One notable omission from the list is that of French AI unicorn Mistral<\/a>, which has positioned itself as an open source alternative<\/a> to incumbents such as OpenAI. While nobody from Mistral responded to TechCrunch for comment, Haji\u010d did confirm that he tried to initiate conversations with the startup, but to no avail.<\/p>\n \u201cI tried to approach them, but it hasn\u2019t resulted in a focused discussion about their participation,\u201d Haji\u010d said.<\/p>\n The project could still gather new participants as part of the EU program that\u2019s providing funding, though it will be limited to EU organizations. This means that entities from the U.K. and Switzerland won\u2019t be able to take part. This flies in contrast to the Horizon R&D program, which the U.K. rejoined in 2023<\/a> after a prolonged Brexit stalemate and which provided funding to HPLT.<\/p>\n The project\u2019s top-line goal, as per its tagline, is to create: \u201cA series of foundation models for transparent AI in Europe.\u201d Additionally, these models should preserve the \u201clinguistic and cultural diversity\u201d of all EU languages \u2014 current and future.<\/p>\n What this translates to in terms of deliverables is still being ironed out, but it will likely mean a core multilingual LLM designed for general-purpose tasks where accuracy is paramount. And then also smaller \u201cquantized\u201d versions, perhaps for edge applications where efficiency and speed are more important. <\/p>\n \u201cThis is something we still have to make a detailed plan about,\u201d Haji\u010d said. \u201cWe want to have it as small but as high-quality as possible. We don\u2019t want to release something which is half-baked, because from the European point-of-view this is high-stakes, with lots of money coming from the European Commission \u2014 public money.\u201d<\/p>\n While the goal is to make the model as proficient as possible in all languages, attaining equality across the board could also be challenging. <\/p>\n \u201cThat is the goal, but how successful we can be with languages with scarce digital resources is the question,\u201d Haji\u010d said. \u201cBut that\u2019s also why we want to have true benchmarks for these languages, and not to be swayed toward benchmarks which are perhaps not representative of the languages and the culture behind them.\u201c<\/p>\n In terms of data, this is where a lot of the work from the HPLT project will prove fruitful, with version 2.0<\/a> of its dataset released four months ago. This dataset was trained 4.5 petabytes of web crawls and more than 20 billion documents, and Haji\u010d said that they will add additional data from Common Crawl<\/a> (an open repository of web-crawled data) to the mix.<\/p>\n In traditional software, the perennial struggle<\/a> between open source and proprietary revolves around the \u201ctrue\u201d meaning of \u201copen source.\u201d This can be resolved by deferring to the formal \u201cdefinition\u201d as per the Open Source Initiative, the industry stewards of what are and aren\u2019t legitimate open source licenses<\/a>. <\/p>\n More recently, the OSI has formed a definition of \u201copen source AI<\/a>,\u201d though not everyone is happy with the outcome. Open source AI proponents argue that not only models should be freely available, but also the datasets, pretrained models, weights \u2014 the full shebang. The OSI\u2019s definition doesn\u2019t make training data mandatory, because it says AI models are often trained on proprietary data or data with redistribution restrictions.<\/p>\n Suffice it to say, the OpenEuroLLM is facing these same quandaries, and despite its intentions to be \u201ctruly open,\u201d it will probably have to make some compromises if it\u2019s to fulfill its \u201cquality\u201d obligations.<\/p>\n \u201cThe goal is to have everything open. Now, of course, there are some limitations,\u201d Haji\u010d said. \u201cWe want to have models of the highest quality possible, and based on the European copyright directive<\/a> we can use anything we can get our hands on. Some of it cannot be redistributed, but some of it can be stored for future inspection.\u201d<\/p>\n What this means is that the OpenEuroLLM project might have to keep some of the training data under wraps, but be made available to auditors upon request \u2014 as required for high-risk AI systems under the terms of the EU AI Act<\/a>.<\/p>\n \u201cWe hope that most of the data [will be open], especially the data coming from the Common Crawl,\u201d Haji\u010d said. \u201cWe would like to have it all completely open, but we will see. In any case, we will have to comply with AI regulations.\u201d<\/p>\n Another criticism that emerged in the aftermath of OpenEuroLLM\u2019s formal unveiling was that a very similar project launched in Europe just a few short months previous. EuroLLM<\/a>, which launched its first model in September<\/a> and a follow-up in December<\/a>, is co-funded by the EU<\/a> alongside a consortium of nine partners. These include academic institutions such as the University of Edinburgh and corporations such as Unbabel, which last year won<\/a> millions of GPU training hours on EU supercomputers.<\/p>\n EuroLLM shares similar goals to its near-namesake: \u201cTo build an open source European Large Language Model that supports 24 Official European Languages, and a few other strategically important languages.\u201d<\/p>\n Andre Martins, head of research at Unbabel, took to social media<\/a> to highlight these similarities<\/a>, noting that OpenEuroLLM is appropriating a name that already exists. \u201cI hope the different communities collaborate openly, share their expertise, and don\u2019t decide to reinvent the wheel every time a new project gets funded,\u201d Martins wrote.<\/p>\n Haji\u010d called the situation \u201cunfortunate,\u201d adding that he hoped they might be able to cooperate, though he stressed that due to the source of its funding in the EU, OpenEuroLLM is restricted in terms of its collaborations with non-EU entities, including U.K. universities. <\/p>\n The arrival of China\u2019s DeepSeek<\/a>, and the cost-to-performance ratio it promises, has given some encouragement that AI initiatives might be able to do far more with much less than initially thought. However, over the past few weeks, many have questioned the true costs<\/a> involved in building DeepSeek.<\/p>\n \u201cWith respect to DeepSeek, we actually know very little about what exactly went into building it,\u201d Peter Sarlin, who is technical co-lead on the OpenEuroLLM project, told TechCrunch.<\/p>\n Regardless, Sarlin reckons OpenEuroLLM will have access to sufficient funding, as it\u2019s mostly to cover people. Indeed, a large chunk of the costs of building AI systems is compute, and that should mostly be covered through its partnership with the EuroHPC centers.<\/p>\n \u201cYou could say that OpenEuroLLM actually has quite a significant budget,\u201d Sarlin said. \u201cEuroHPC has invested billions in AI and compute infrastructure, and have committed billions more into expanding that in the coming few years.\u201d<\/p>\n It\u2019s also worth noting that the OpenEuroLLM project isn\u2019t building toward a consumer- or enterprise-grade product. It\u2019s purely about the models, and this is why Sarlin reckons the budget it has should be ample. <\/p>\n \u201cThe intent here isn\u2019t to build a chatbot or an AI assistant \u2014 that would be a product initiative requiring a lot of effort, and that\u2019s what ChatGPT did so well,\u201d Sarlin said. \u201cWhat we\u2019re contributing is an open source foundation model that functions as the AI infrastructure for companies in Europe to build upon. We know what it takes to build models, it\u2019s not something you need billions for.\u201d<\/p>\n Since 2017, Sarlin has spearheaded AI lab Silo AI, which launched \u2014 in partnership with others, including the HPLT project \u2014 the family of Poro<\/a> and Viking open models<\/a>. These already support a handful of European languages, but the company is now readying the next iteration \u201cEuropa\u201d models, which will cover all European languages.<\/p>\n And this ties in with the whole \u201cnot starting from scratch\u201d notion espoused by Haji\u010d \u2014 there is already a bedrock of expertise and technology in place.<\/p>\n As critics have noted, OpenEuroLLM does have a lot of moving parts \u2014 which Haji\u010d acknowledges, albeit with a positive outlook.<\/p>\n \u201cI\u2019ve been involved in many collaborative projects, and I believe it has its advantages versus a single company,\u201d he said. \u201cOf course they\u2019ve done great things at the likes of OpenAI to Mistral, but I hope that the combination of academic expertise and the companies\u2019 focus could bring something new.\u201d<\/p>\n And in many ways, it\u2019s not about trying to outmaneuver Big Tech or billion-dollar AI startups; the ultimate goal is digital sovereignty: (mostly) open foundation LLMs built by, and for, Europe.<\/p>\n \u201cI hope this won\u2019t be the case, but if, in the end, we are not the number one model, and we have a \u2018good\u2019 model, then we will still have a model with all the components based in Europe,\u201d Haji\u010d said. \u201cThis will be a positive result.\u201d<\/p>\n<\/div>\nUp to scratch<\/h2>\n
Build up<\/h2>\n
The open source definition<\/h2>\n
Two for one<\/h2>\n
Funding gap<\/h2>\n
Sovereign state<\/h2>\n