AI Models

An AI model is a trained neural network that has learned a pattern from examples — and uses that pattern to process new inputs. The word “model” is meant literally: a mathematical replica of what was contained in the training data. When you chat with ChatGPT, Claude or Gemini today, you are talking to a language model — currently the most prominent form. But there are also models for images, audio, video and code, and increasingly ones that handle several of these modalities at the same time.

To draw a clean line: “AI” is the umbrella term for everything that makes machines appear intelligent. An “AI model” is the specific piece of software that does the actual work. And an “algorithm” is the method used to train the model — not the model itself.

How an AI model works in broad terms

During training, the model is fed huge amounts of data — for large language models (LLMs) that means texts from the web, books, code repositories and much more. From this data, it learns probabilities: which word typically follows which? Which image regions belong together? The result is a set of parameters — numbers that store the learned pattern. A modern language model has billions to trillions of them.

When you later ask the model a question, it splits your text into tokens (small word or character chunks) and predicts the next most likely token step by step — and then the one after that, and so on, until the answer is complete. That is the whole trick: predicting the next token, billions of times per response.

What a model is good at depends on three factors: the training data, the model size (more parameters = often more capability, but also more expensive) and post-training — for example fine-tuning on specialised tasks or RLHF, which teaches models to give helpful rather than just probable answers.

Which AI models exist?

Today’s AI models can be roughly split into three categories:

Closed-source models come from commercial providers like OpenAI (GPT series), Anthropic (Claude), Google (Gemini) or xAI (Grok). They run on the providers’ servers and are accessible via APIs or chat apps. Strengths: usually leading on quality and multimodality. Drawbacks: ongoing per-request costs, data protection questions, vendor lock-in.

Open-weight models like Meta Llama, Mistral, DeepSeek, Alibaba Qwen or Z.ai GLM can be downloaded and run yourself — on your own servers or via specialised hosters. The gap to closed models has shrunk significantly over recent years; for many tasks, open-weight models are now competitive or better.

Local models are a sub-category: smaller or heavily compressed open-weight models that run on a normal computer or laptop — via tools like Ollama or LM Studio. Advantages: free, offline, privacy-friendly. Limits: less capable than the big cloud models, need enough RAM and ideally a GPU.

Within each category there are additional specialisations: coding models (GPT-Codex, Codestral, DeepSeek-Coder), reasoning models for complex problem-solving, multimodal models for image/audio/video, and small fast variants (“Mini”, “Flash”, “Haiku”) for high volume at low cost.

What you’ll find on this page

Below you’ll see a topic world around AI models: current news on new releases, blog articles with background and practice, lexicon articles for deeper dives and a glossary of the most important terms. Use the filters above to switch by provider — OpenAI, Anthropic, Google, Meta, Mistral, DeepSeek and more — or look at the topic LLM Pricing if you’re mainly interested in costs and billing models.