Open Source

Open does not mean simple.

Open AI can be brilliant. It can also be confusing. "Open source", "open weights", "free app", "research release" and "local model" are not the same thing.

A dark maker-space desk with model blocks, code cards and a glowing shared blueprint
Plain definition

There are four levels of open.

When someone says an AI is open, ask what part is open. The answer changes whether you can trust it, inspect it, improve it, sell with it or run it locally.

Open source

Best

The code is available under a recognised open-source licence. You can inspect, modify and share it under the licence rules.

Open weights

Common

The trained model files are available, but the training data and full recipe may not be. Many "open AI" releases are really open-weight releases.

Open API

Useful

You can call the model through a service, but the model itself stays closed. You are renting access.

Free app

Not open

A free app may still be closed, data-hungry or heavily limited. Free is a price, not a licence.

Projects to know

Useful open and open-weight shelves.

These are not beginner recommendations by themselves. They are places or model families worth understanding.

Meta Llama

62

Widely used open-weight model family. Good for builders, but the licence has its own rules and is not the same as pure open source.

Llama site

Mistral open models

72

European open-weight releases alongside commercial products. Useful when you want a non-US alternative and strong model performance.

Mistral

DeepSeek

52

Efficient open-weight reasoning models. Technically important, but beginners should pay close attention to data handling and jurisdiction.

DeepSeek-R1

Alibaba Qwen

60

Broad open-weight model family with strong text, code and multimodal releases. Check each licence and deployment path.

Qwen GitHub

Hugging Face

77

The library shelf for models, datasets, demos and documentation. Trust depends on the specific model, not the hub name alone.

Hugging Face

Ollama and local runners

Practical

Local runners can help keep prompts on your machine, but you still need the right model, enough hardware and a licence you can live with.

Ollama
Local AI

Running locally is control, not magic.

A local model can keep prompts on your computer if set up properly. It can still be wrong, biased, insecure, too slow or unsuitable for private work if you do not understand what it is.

Good local uses

Private drafting, coding practice, offline experiments and learning how models behave.

Beginner traps

Large downloads, confusing licences, weak laptops, unsafe outputs and old models that cannot handle your task.

Check first

Model licence, hardware needs, community reputation, safety notes and whether uploads leave your computer.