The leading providers in the field of language models are currently:

• OpenAI (USA): Developer of GPT-4o, o1, and o3. Best-known product: ChatGPT.
• Anthropic (USA): Developer of Claude. Focus on security and reliability.
• Google DeepMind (USA/UK): Developer of Gemini. Deeply integrated into Google products.
• Meta (USA): Developer of Llama. Open-source models that can be run independently.
• Mistral (France): European provider, strong open-source models.
• DeepSeek (China): Very powerful models developed with significantly lower budgets.

The landscape is changing rapidly. At nuwai, we continuously evaluate which model is best suited for which application.

GPT-4 (or GPT-4o, GPT-4 Turbo) is the model – meaning the underlying technology. ChatGPT is the product – the user interface through which you interact with the model.

You can imagine it this way: GPT-4 is the engine, ChatGPT is the car. Other products can use the same engine – via the so-called API. This is exactly how we at nuwai build customized solutions: We use the best available models and build the appropriate application around them.

API stands for “Application Programming Interface” – an interface through which software can communicate with other software.

In the AI context, this means: Instead of opening ChatGPT in a browser, companies can integrate AI models directly into their own systems – into their CRM, their website, their internal toolchain. Via the API, a text (prompt) is sent, and the response is received back, automated and without manual input. This is the basis for any professional AI integration.

Closed Source means that the code and model weights remain secret. You can only use the model via the provider’s API. Examples: GPT-4, Claude, Gemini.

Open Source (or Open Weight) means that the model weights are publicly available. You can download the model and run it on your own infrastructure – without dependency on the provider. Examples: Llama (Meta), Mistral, DeepSeek.

For companies with strict data protection requirements, an open-source model on their own hardware may be the best option.

Fine-tuning is the process of further adapting an already trained model with additional, specific data. This means taking a “generalist” model and specializing it for a particular task or industry.

Example: A language model is retrained with thousands of customer service conversations from an insurance company so that it masters the tone, products, and typical questions of that company particularly well. Fine-tuning is more complex than pure prompt engineering but often delivers better results for specialized applications.

RAG is a method where a language model not only responds based on its training but first searches an external knowledge base for relevant information and incorporates it into the answer.

Imagine giving an employee a company manual and saying: “Answer customer questions, but first consult the manual.” RAG does exactly that – only automated. It drastically reduces hallucinations and makes it possible to connect AI with company-specific, current knowledge without retraining the entire model.

Embeddings are numerical representations of text (or images) in a high-dimensional space. Simply put: A sentence is converted into a long list of numbers that capture its “meaning.”

Why is this useful? Because it allows you to calculate semantic similarity. “Dog” and “puppy” are close in the embedding space, while “dog” and “tax return” are far apart. Embeddings are the basis for RAG systems, semantic search, and recommendation systems.

A vector database stores embeddings and allows for quick searches for similar content. Instead of searching for exact matches like a traditional database (“show all entries with the word car”), it searches for semantic similarity (“show everything thematically related to mobility”).

Well-known vector databases include Pinecone, Weaviate, Qdrant, and Chroma. They are a core component of modern AI applications – especially when it comes to making company-specific knowledge accessible to AI systems.