Local and private
What are open-weight models?
Open-weight models publish their trained parameters so you can run them yourself. What "open" covers, what it does not, and why licences differ so much.
All explainers · Last reviewed:
What "open" covers here
Open weights means the numbers that constitute the trained model are downloadable. That is enough to run it, fine-tune it, inspect it, and keep it working regardless of what the publisher does later.
It usually does not include the training data, and often does not include the training code. So an open-weight model is reproducible in the sense that you can run it, not in the sense that you could rebuild it.
The licences genuinely differ
Some open-weight models carry ordinary permissive licences. Others carry conditions: restrictions on commercial use above a size threshold, prohibitions on particular applications, or requirements about attribution and derived models.
This matters commercially and is easy to skip. "We can download it" and "we can use it in our product" are different questions, and only the licence answers the second.
Why they matter
They are the only models you can run entirely on your own hardware, which makes them the foundation of every local and private deployment. They also cannot be retired underneath you — a downloaded model works for as long as you keep it.
The capability gap to the best hosted models is real and has narrowed considerably. For a large share of everyday work it is no longer the deciding factor.
Questions people ask
- Is open-weight the same as open source?
- No. Open source implies the source and the freedom to use and modify it. Open weight means the parameters are published, under whatever licence the publisher chose — which is sometimes restrictive.
- Can I fine-tune an open-weight model?
- Technically yes, that is one of the main reasons to want the weights. Whether you may, and what you may do with the result, is a licence question that varies by model.
- Are they safe to use commercially?
- Many are; some are not without conditions. Read the specific licence for the specific model — this is the one thing in this area that genuinely cannot be generalised.
Try it rather than read about it
ClawAI runs open-weight models through Ollama and llama.cpp on your own hardware, alongside 5 cloud providers, with routing deciding which handles what.