HELSINKI — In a strategic collaboration with TurkuNLP from the University of Turku and the EU-funded HPLT project, Silo AI has unveiled the Viking 33B, a groundbreaking multilingual AI model designed to strengthen the AI landscape in Europe. This new model, which is powered by AMD’s computing platforms, is tailored for low-resource languages and significantly advances the capabilities of large-scale language models (LLMs) for Nordic languages, including Danish, Finnish, Norwegian, Icelandic, Swedish, and various programming languages.

The Viking 33B model is the culmination of a series of successful releases, including Viking 7B, Viking 13B, and Poro 34B. This 33 billion parameter model has been developed with a highly optimized dataset of 2 trillion tokens, providing extensive language processing capabilities. It also supports translation between English and Nordic languages, making it a powerful tool for enterprises and researchers alike.

The model’s architecture leverages advanced features like flash attention, rotary embeddings, and grouped query attention, following a structure similar to Llama 2. Designed to operate on LUMI, Europe’s fastest supercomputer, the Viking 33B model uses up to 1024 AMD MI250X GPUs, demonstrating the feasibility of training large-scale LLMs with high-performance computing.

Silo AI and TurkuNLP are committed to making Viking 33B an open-source model under the Apache 2.0 License, ensuring accessibility for both commercial and research applications. The collaboration between the two organizations aims to bring diversity to AI models and accelerate the integration of AI into various industries. By working together, they aim to increase Europe’s digital sovereignty and boost the continent’s technological competitiveness on the global stage.

Peter Sarlin, CEO and co-founder of Silo AI, emphasized the strategic importance of the project, saying, “This new release represents a significant step forward in our mission to provide cutting-edge, open-source AI models. By integrating the capabilities of AMD platforms with the linguistic strengths of Viking 33B, we’re creating opportunities for enterprises to drive AI adoption and scale innovation.”

The Viking 33B model also serves as a major advancement in AI’s role in Europe, building on the strong partnerships that Silo AI has forged with various research and corporate entities. By utilizing AMD’s high-performance GPUs, the model is designed to handle the most demanding AI workloads and is optimized for efficient training and inference.