Run your own AI · Open weight vs open source
Open weight, open source and the rest
thecontinuitylayer.com/open/learn/open-weight-vs-open-sourceChecked 6 November 2026. US dollars. No affiliate links.
“Open” gets used for eight different things in AI. They are not interchangeable, and the difference decides what you can do with a model: run it, change it, sell something built on it, or see how it was made.
By the end
You can open any model's page, find its license, and say which of these eight it is.
The eight terms
| Term | What it means | Example |
|---|---|---|
| Open weights | You can download the trained model file and run it yourself. Says nothing about how it was trained or what the license allows. | Llama 3.1, Qwen3.8, DeepSeek-V4.1-Flash |
| Open source | The weights, the training code and the training data are all published, under a license that lets anyone use them. Rare for models. | Ai2's Olmo: data, code and checkpoints released |
| Permissive license | Use, change and sell with almost no conditions beyond keeping the notices. Apache 2.0 and MIT are the common ones. | Gemma 4, gpt-oss, Phi-4, DeepSeek-V4.1-Flash |
| Custom (restricted) license | Weights are downloadable, but the license adds rules: a usage policy, a required credit line, or a revenue cap above which you need the maker's permission. | Llama 3.x (700M-user clause), Muse Spark Open, Kimi K3 |
| Community retune | Someone other than the maker fine-tuned the model, often to follow the user's instructions over the maker's defaults. The original license still applies. | Dolphin 3, Hermes 3 (both built on Llama 3.1) |
| Local | The model runs on your own computer. Nothing you type has to leave it. | Ollama or LM Studio on a laptop |
| Self-hosted cloud | You run the model on a server you rent. You control the setup; the server company controls the machine. | A rented GPU on Runpod or Lambda |
| Fully independent | Local or self-hosted, open-source software on top, your own data store, no account with any AI company. Possible today, and real work to keep running. | Ollama + Open WebUI on your own box |
Read a license in two minutes
Open the model's page on Hugging Face. The license is in the panel on the right, or in a file called LICENSE. Then ask three questions:
- Is it Apache 2.0 or MIT? Then you can use it for almost anything, including selling a product built on it.
- Is it a custom license? Search the text for “policy”, “revenue”, “users” and “attribution”. Those are where the conditions live.
- Is the download gated? Some makers ask for your name and date of birth before you can download. That is a condition too.
Every model in the model finder links its license and says, in one line, what rules it adds.