Skip to main content

Open release of the first large Nordic language model GPT-SW3

Thursday, November 16, 2023

AI Sweden now releases the first large Nordic language model, GPT-SW3. It is available as an open model for businesses and organizations to use in their products and services. The model provides Sweden with a crucial resource that strengthens our ability to use AI for increased competitiveness and a stronger welfare system. Developed by AI Sweden in collaboration with WASP and RISE, GPT-SW3 is specifically developed for Swedish and Nordic languages.

Scrabble tiles tumbling in the air, each spelling out the letters and symbols of GPT-SW3

The launch of GPT-SW3 interpreted by OpenAI's DALL-E 3

GPT-SW3 can be used for a variety of tasks such as building solutions for text analysis, classification, text generation, text processing, content moderation, and chat functionalities.

As a result, it addresses some of the challenges faced by both the private sector and the public sector concerning other available large-scale language models. Some of these obstacles include ambiguities surrounding training data, unclear licensing terms, or that models are only accessible as cloud services.

There is a significant need for a model trained on Swedish text with clear licensing terms that can be fine-tuned to meet specific requirements and run on dedicated hardware, both for legal reasons and business competitiveness. GPT-SW3 fulfills all these needs, thus opening the doors to a multitude of new applications within Swedish businesses and the public sector.

We have noticed tremendous interest in utilizing large language models among Swedish organizations. There are many ideas on how this technology can be applied but not all have been feasible with existing models. With GPT-SW3, we now have a new resource for strengthening Sweden’s competitiveness and our society.

Magnus Sahlgren, Head of Research NLU at AI Sweden and leader of the GPT-SW3 project

Various studies on generative AI consistently highlight significant productivity gains. A 2019 report by DIGG estimated that AI could contribute an annual value of 140 billion SEK to the public sector. This estimation was made before generative AI gained widespread adoption.

We view GPT-SW3 as a pivotal piece of the puzzle in unlocking the immense value that can be created by using AI. Studies have already demonstrated how GPT models can streamline or improve the quality of various tasks. With a Swedish-developed model focused on the Swedish language, the possibilities for the public sector, in particular, to reap these benefits are greatly increased.

Martin Svensson, Managing Director at AI Sweden.

Many organizations in both the private and public sectors are already working with generative AI, including language models. For example, Region Halland, explores how language models can help to summarize medical records, predict adverse events, code journals, extract information from free text, and create discharge messages for patients.

AI Sweden's language model GPT-SW3 opens up new opportunities, and there is great value in having full control over the model itself. It is evident that language models can significantly support the healthcare sector. We now look forward to validating critical applications in clinical settings

Markus Lingman, senior physician, professor, and chief strategy officer at Halland hospital group.

This is GPT-SW3

GPT-SW3 is a generative pre-trained model. By thoroughly analyzing vast amounts of text, the language model discovers intricate language patterns. These patterns can then be leveraged to intelligently generate text: When you input a shorter or longer text into a GPT model, it will seamlessly generate statistically plausible continuations.

However, it's important to note that GPT-SW3 is not a ready-to-use product or service. To harness its capabilities, developers must build something using GPT-SW3 as their foundation. 

Developers can access different versions of GPT-SW3 on Hugging Face, along with the license and terms of use.

GPT-SW3 is the result of a collaboration between AI Sweden, WASP, and RISE with funding from Vinnova.

Read more