Meet the AI Model Ecosystem
Get a map of who does what in the AI world: model makers, hubs, hosts, tools and app builders.
Before you can choose a model you need a rough map of where models come from, because the AI world has its own supply chain, and new names appear every month.
Model makers train the models. A handful of big labs, including OpenAI, Anthropic, Google, Meta, Mistral AI and DeepSeek, spend months and fortunes on pre-training and fine-tuning (Track 2). Some keep their models closed and only offer them through their own API. Others publish the weights for anyone to download: more on that difference next lesson.
Model hubs are the libraries for those open models. Hugging Face’s hub is the best known, with hundreds of thousands of models: big general ones, small specialised ones, fine-tuned and quantized variants made by the community.
Hosts and inference providers run models for you behind an API: the makers’ own services, big clouds such as Amazon Bedrock, Google Cloud or Microsoft Azure that offer many companies’ models under one roof, and specialists such as Together AI or Groq that focus on serving open models quickly and cheaply, using the tricks from Track 4.
Run-it-yourself tools let you host a model on your own hardware: serving engines like vLLM for servers, and friendly apps like Ollama for a laptop.
App-building tools sit on top: libraries that help apps send prompts, connect tools and documents, and switch between providers.
The lines blur. Google makes models and hosts them; Hugging Face is a hub that also hosts. But the map tells you which questions to ask about any new name: does it make models, store them, run them, or help you use them?
Key takeaways
- Model makers train models; some keep them closed, others publish the weights.
- Hubs such as Hugging Face store and share open models.
- Hosts and inference providers run models behind an API; clouds offer many makers’ models in one place.
- Serving engines and local apps let you run models yourself; libraries help apps use them.
Check your understanding
Go deeper (more technical)
The ecosystem moves fast: model rankings change month to month, and new providers appear constantly. A healthy habit is to design apps so the model and provider can be swapped, for example by keeping prompts and tests separate from provider-specific code.
Model cards (short documents published with a model) list what it was trained for, its context length, licence and known limitations. They’re the first thing to read before trying a new model.
Company names are examples of each layer, not recommendations.