How generative models write, reason, see and build — and the language models leading the way.
Generative AI describes models that produce new content — text, images, audio, video and code — rather than only classifying what already exists. Trained on very large corpora, they learn the statistical structure of language and media well enough to continue, translate, summarise and invent. The result is a general-purpose interface: you describe the outcome you want in plain language, and the model drafts it.
The transformer's self-attention mechanism lets a model weigh every token against every other, which is what makes long-range reasoning over text practical. Scaled up, these become the large language models behind chat assistants and coding tools.
Read moreImage, audio and video generators learn to reverse a noising process: starting from random noise, they denoise step by step toward a sample that matches your prompt. This underpins tools such as Stable Diffusion, Imagen and modern video models.
Read moreModern frontier models accept more than text. Images, documents, audio and screen recordings are encoded into the same representation space as language, so a single model can read a chart, describe a photo or hold a spoken conversation.
Read moreEmbeddings turn content into vectors that capture meaning, enabling semantic search. Retrieval augmented generation puts the retrieved passages into the prompt so answers are grounded in your own documents instead of the model's memory alone.
Read morePre-training is followed by instruction tuning and preference optimisation — RLHF, DPO and related methods — which teach a model to follow instructions, refuse unsafe requests and adopt a consistent voice. LoRA and other adapters make this affordable.
Read moreGiven tools — search, a shell, a database, an API — a model can plan, act, observe the result and iterate. Function calling and protocols such as MCP standardise how those tools are exposed, turning a chat model into a system that gets work done.
Read moreA snapshot of widely used model families. Availability and versions change quickly — check each provider for current releases.
| Model family | Developer | Access | Known for |
|---|---|---|---|
| GPT | OpenAI | Proprietary API | General reasoning, multimodal chat, broad tooling ecosystem |
| Claude | Anthropic | Proprietary API | Long-context work, coding and agentic workflows |
| Gemini | Google DeepMind | Proprietary API | Very long context and native multimodal input |
| Llama | Meta | Open weights | Self-hosting and fine-tuning at many parameter sizes |
| Mistral & Mixtral | Mistral AI | Open weights + API | Efficient dense and mixture-of-experts models |
| DeepSeek | DeepSeek AI | Open weights + API | Strong reasoning and code at low inference cost |
| Qwen | Alibaba Cloud | Open weights + API | Multilingual coverage and a wide size range |
| Grok | xAI | Proprietary API | Real-time context from the X platform |
| Command | Cohere | Proprietary API | Enterprise retrieval and RAG-oriented deployments |
| Gemma & Phi | Google / Microsoft | Open weights | Small models that run on a laptop or at the edge |
Pick a hosted API for the fastest path, or open weights when you need control over data and cost. Start with a clear prompt, add retrieval when answers must cite your own sources, and measure quality with a small evaluation set before you scale.