The honest version: most "open source AI" you hear about is open weights, not open source by the strict definition the Open Source Initiative defends. The weights are downloadable, but the training data and full recipe usually are not, so you cannot fully reproduce or audit the model. Even OpenAI, despite the name, kept its flagship closed until releasing separate open-weight gpt-oss models in 2025.
The distinction matters for trust and control, not marketing. A truly open model can be inspected and self-hosted; an open-weights model can be run and fine-tuned but not fully verified. For brand visibility, the practical effect is that strong open models (DeepSeek, Qwen, Kimi) are now everywhere, so the engines and apps citing you increasingly run on them.