Large Language Models (LLMs)

Large Language Models (LLMs) are very large neural language models with billions of parameters, typically based on the Transformer architecture described by Vaswani et al. (2023)¹ and pretrained on large text corpora.^{2 3 4} Due to model size and the diversity of training data, LLMs develop so-called emergent abilities - new capabilities that were not explicitly programmed and only appear at a certain scale.⁵

Such foundation models serve as a universal basis for a wide range of downstream tasks. Well-known examples such as GPT-4o^{3 6} demonstrate the potential of such systems: they act as powerful general-purpose language assistants that understand complex instructions, draw multi-step logical conclusions, and solve a variety of tasks.⁴ The ability to generalize across many contexts makes LLMs central building blocks for general AI applications.^{4 5}

Below, LLMs are introduced and analyzed with respect to their key characteristics, application areas, and technological orientations. The classification distinguishes between open-source and proprietary language models, taking into account key differences in transparency, licensing models, access options, and infrastructure requirements.

Comparison of Open-Source and Proprietary LLMs

With the growing adoption and specialization of large language models, a systematic classification of their properties and deployment potential becomes essential. An overview of selected models highlights key differences.

Open-Source Language Models

Open-source models offer decisive advantages for research and development through their transparency and adaptability. Models from Meta, Mistral AI, Alibaba Cloud, and DeepSeek demonstrate how even compact architectures can compete with much larger systems.

Proprietary Language Models

Proprietary models are characterized by closed architectures and restricted access, but offer high scalability, performance, and integration into production systems.

Conclusion

Large language models are the foundation of modern AI development. They enable both specialized and universal applications and develop new emergent abilities through scaling. Whether open or proprietary, LLMs are shaping the future of human-machine interaction.