A side project to test what a small team could achieve by fine-tuning and training language models specifically for Swedish.
At Incredible, we spend much of our time exploring how artificial intelligence can be applied to real-world problems. As part of that work, we launched SWE-LLM: a side project created to test what a small team could achieve by fine-tuning and training language models specifically for Swedish.
The project began with a simple question: could we build a model that produced high-quality Swedish text, despite working with significantly fewer resources than the world's largest AI companies?
The result was SWE-LLM, a Swedish-language model designed to generate natural, grammatically correct, and stylistically consistent Swedish content.
"Building SWE-LLM was an opportunity for us to better understand the entire model-development process, from preparing training data to fine-tuning, evaluating, and deploying the finished model. It also demonstrated how much a focused team can accomplish with limited resources."
— Philip Alm, co-founder of Incredible
SWE-LLM is based on Llama 3 70B and was fine-tuned using a carefully curated dataset containing more than 12,700 Swedish texts and approximately 13.3 million tokens.
The model was trained using NVIDIA A100 GPUs and was made publicly available so that researchers, developers, companies, and other interested users could experiment with it.
SWE-LLM was trained to write in standard Swedish, or rikssvenska. It can generate text with strong grammar, structure, and style, with particular strengths in areas such as articles, technical reports, explanations, and other forms of long-form content.
The training data covered more than 250 types of text, ranging from news-style articles and professional communication to technical and creative writing.
By including conversational examples in the training process, we also explored how the model could perform in chat-based applications and respond naturally to instructions written in Swedish.
The project helped us evaluate how specialised Swedish-language models could support use cases such as:
"SWE-LLM gave us practical experience working with Swedish training data and evaluating model behaviour in our own language. It allowed us to explore tasks directly in Swedish that are often designed, tested, and optimised primarily in English."
— David Bacelj, co-founder of Incredible
SWE-LLM was never intended to compete with the largest general-purpose models on every benchmark. Instead, it was an experiment in specialisation.
Large international models are highly capable, but smaller or more focused models can still offer advantages when they are adapted to a particular language, industry, workflow, or type of task. Through the project, we wanted to explore how carefully selected training data could influence linguistic quality, tone, and domain-specific performance.
It also gave our team hands-on experience with the practical challenges involved in model development, including:
These lessons have continued to inform how we approach the development of AI-powered products and automated workflows at Incredible.
SWE-LLM generating a story about the Swedish author Astrid Lindgren.
Our primary focus at Incredible is building useful AI products and systems that solve practical problems. However, experimental projects such as SWE-LLM allow us to investigate the underlying technology more deeply.
One area we are particularly interested in is the development of AI systems that can interact with existing applications and participate in larger workflows. For AI to create meaningful business value, models need to do more than generate text. They need to understand instructions, make decisions, use tools, and communicate reliably with other systems.
"We believe it is important to understand how models are built, not only how to integrate them. SWE-LLM was one step in developing that understanding. The experience has helped us make better decisions when designing AI agents, applications, and automation systems."
— Philip Alm, co-founder of Incredible
Incredible is a technology company based in Jönköping, Sweden. We build AI-powered products and systems designed to simplify work, automate processes, and help organisations make practical use of large language models.
Our work combines product development, applied artificial intelligence, workflow automation, and experimentation with emerging AI technologies.
SWE-LLM represents one of those experiments: a focused project that allowed us to explore what it takes to build, train, and deploy a Swedish-language model ourselves.
Researchers, developers, and organisations interested in the project are encouraged to explore SWE-LLM and experiment with its capabilities.
For questions about SWE-LLM, Incredible, or potential collaborations, please contact the Incredible team through our current website or company contact channels.
Get started on macOS