Quick correction that saves people real money: open weight is not open source. People use them like they are. They are not. And the distinction bites exactly when it matters most, when you ship something.
Open weight means the trained weights are published. Download them, run them. Generous and useful already. But the license on those weights decides what you can do with them. Some sit under permissive licenses, Apache 2.0, MIT, close to true open source. Others carry community licenses with conditions: usage thresholds, restrictions on certain applications, keep-the-license-attached requirements. Meta's Llama licenses are the famous example. Read the actual terms. Not the summary. The terms.
Then there is everything not published. Training data, training code, full reproducibility. Truly open source AI would include all of it. Almost no major model does. So "open weight" is the honest label for most of what we call open models, and our database uses it on purpose: 200 open-weight LLMs, license spelled out per model. When the label and the license disagree, the license wins. Every time.
Why should you care? "Can I run it" and "can I ship a product on it" are different questions with different answers. Answer the second one before you build. For a business, it shapes the whole private AI setup for business plan: which models your on-premise AI deployment can use, which ones the lawyers bless, which ones stay in the lab.
Our open weight LLM models list tracks licenses next to specs and benchmarks, with an LLM license comparison for commercial use built in, so the question is answerable in one place. Ready to run it all privately? I do private AI setup for businesses at privateaiagent.fyi. Your data never leaves.