The Institute for Language and Speech Processing (ILP) of the Athena Research Center – the only institute in Greece which developed machines capable of answering questions and solving problems by imitating human intelligence – has just launched its second model of Large language (LLM), Llama -Krikri.
Llama-Krikri follows the traces of the institute’s mother tongue model, Meltemi. Although he cannot compete with Deepseek or Chatgpt, he performs the same fundamental tasks. A conversation with the team behind it looks like a deep dive into the fascinating world of artificial neural networks, which treat and generate human languages.
“ It is as defining the language model to pass through your stored data and use it as “cheat sheet” to generate responses’
Why the name Llama-Krikri? “It is an association game game,” explains Vassilis Katsouros, director of ILA at Athena Research Center, with a smile. “The name comes from the basic model we have used – the Meta Llama 3.1 3.1 model – combined with the Greek equivalent of Lama, the Kri -Kri (a type of wild goat from Crete).”
Llama-Krikri is much lower in calculation scale than Deepseek, but how much? “If the Deepseek neural network has 671 billion parameters, 37 billion of which are activated according to the user’s request, our model has only 8 billion parameters,” explains Katsouros. To put it simply, by translating the Greek language into numbers – since the numbers of computers’ process, not the words – llama -krikri represents it using 8 billion digital values. “It’s a different investment scale,” he adds. “High -end language models require large amounts of data, immense calculation power and significant financial resources.”
Digital aid
Llama -Krikri accomplishes everything Meltemi has done – and does a much better work. What exactly does Meltemi do? It allowed Greek school, research and business communities to build their own digital assistants. “Although it does not provide a service to respond to direct requests from the general public, it allows anyone with knowledge in programming them to their needs, to submit queries and to receive responses according to their own Data, ”explains Sokratis Sofianopoulos, who, with Prokopis Prokopidis, Georgios paraskevopoulos, Leon Voukoutis and Dimitris Roussis – all scientific partners in ILS – have developed Llama -Krikri.
“It is as defining the linguistic model to pass through your stored data and use it as” cheat sheet “to generate answers,” he adds. And these responses arise in modern fluid Greek as well as in English. “Meltemi had a huge impact – tens of thousands of downloads on Hugging Face, a platform where the developers of AI share the code. It was downloaded by students and academics, used in countless applications and research projects, stimulated research activity in Greece and increased productivity in small and medium -sized enterprises. He offered an affordable solution for the rapid execution of the project, ”explains Sofianopoulos. In addition, the Institute supported public organizations and businesses in the construction of their own digital assistants. “As part of the European Hubs of Digital Innovation (EDIHS), we work with companies that require personalized adaptations of our linguistic model to meet their specific needs,” adds Katsouros.
So, what improvements does Llama-Krikri bring? “He has a much stronger memory – he does not lose the track during long conversations. While the previous model could remember 8,000 words and maintain a dialogue for only a few minutes, Llama -Krikri can remember up to 128,000 words – a whole book of 250 pages – and can manage a sequence of multiple questions without Forget what was said at the beginning. He can deal with larger volumes of text to generate his answers. It produces better quality text; He writes in the polytonic system, can deal in the old Greek and results in modern Greek – which the previous model has not been formed. He also writes better code, because models are not limited to natural language – they also write computer languages. In addition, it more precisely solves mathematical problems and has stronger reasoning. In other words, it can perform complex tasks that require several intermediate stages of logical thought, ”explains Katsouros.
At the same time, the research team creates and publishes a multitude of linguistic resources, including data sets that reflect data in English in Greek. This effort allows an in -depth evaluation of the performance of the Greek language model, and all this is made openly available. As Maria Giackou said, scientific partner at the Institute, “Meltemi and Llama-Krikri are open source models, which means that any researcher or professional in the industry can freely examine and personalize their operation. It is the power of open source. On the other hand, the Openai model is closed – you get what it gives you, without the possibility of adapting your specific needs. “”
Katsouros also explains that the Greek ecosystem of small and medium -sized enterprises “needs a technology in Greek which meets its unique requirements and its financial capacities”. In addition, like adding Sofianopoulos, a Greek company does not need chatgpt to support its services. It needs a smaller model – or perhaps even a second or a third – to allocate specific tasks, all managed internally to avoid the risk of data leaks and significant costs charged by Openai. This is particularly important when the services are provided by subscription, which is essential to ensure safety. In these models, the cost is based on the number of segments of words generated by each request and response. The more a word segment is broken, the higher the cost. OPENAI and Anthropic often decompose each Greek word into several sub -words – sometimes even the character by character – thus generating Greek text by bringing these fragments together. “In our approach, we break down the Greek words into the greatest possible segments, considerably reducing the calculation cost. On average, we divided a Greek word into 1.5 segments, while other models divide them into 2.5, 6 or even 7 segments, “he explains.
Cultural regime
But these are not the only reasons why we need Greek language models. “Our language carries our culture and shapes the way we think and communicate in Greek. You need a service which, when you ask a culturally specific question, say: “Recommend a healthy breakfast”, replies in an appropriate way rather than suggesting something like “beans with bacon”, which is a British breakfast concept. It is crucial to form large language models – and all linguistic technologies – on texts originally written in Greek rather than on translations of other languages, ”explains Giackou. “It is also a question of digital survival for our language. We know that unwritten languages can disappear; Today, languages without digital presence will also disappear. For a survive language, it must have a digital imprint, which means being supported by linguistic technologies, including major language models. And because we are a very small market, large companies show little commercial interest in our language. If we do not invest in Greek linguistic technologies, no one else will. »»
European deep?
Could Europe develop a deep deep? Katsouros replies: “It could certainly. Deepseek used calculation techniques that are well known in the research community. Similar methods have been adopted by the French company Mistral AI, one of the main European companies in the development of the linguistic model. What is necessary is more experimentation, continuing education, access to resources, hard work and constant adaptation to the latest technological advances.
With the open source models, the effort is shared – the French rely on the work of the Chinese, the Greeks rely on the French, etc. – So a major European language model could indeed emerge and have a significant impact. However, Europe must remain actively involved and develop a more balanced regulatory framework for data copyright – perhaps allowing free use of 10% or 20%, similar to the allowances designed for educational purposes. (French President Emmanuel) Macron observed once in Europe, we regulate too much and innovate too little. »»
“Currently, a European alliance for linguistic technologies has been formed, with 17 countries participating so far. The objective is to collect linguistic data and develop new models of large -scale language. In addition, two major European projects have been funded to create a multilingual model which will also support all official European languages, in accordance with Europe’s multilingual policy, ”explains Giackou.
Sofianopoulos is less optimistic, reminding us that Europe has not yet developed entirely new architectures for large -language models – at least not on a scale that can compete with the initiatives of the United States and China. An initiative concerning a part of this challenge is Pharos, “the Greek factory of AI”, a project involving several domestic and university research organizations. The key partners include National Research and Technology Infrastructures (GRNET), the Demokritos Center for Scientific Research, the National Technical University of Athens (NTUA), Athena RC and the Hellenic Corporation of assets and participations. Pharos will operate using calculation resources from the high performance computer Daedalus (HPC).
“The more people who get involved, the more innovative results we will quickly see,” concludes Katsouros.