OSS

In traditional software, open source means you can read and modify the source code. In AI, the conversation is more nuanced. 'Open weights' means the trained model parameters are publicly downloadable, which lets you run the model locally or fine-tune it. 'Open source' in a fuller sense also includes the training code, data, and methodology. Very few frontier models are fully open source in that sense.

Models like Meta's Llama family, Mistral's models, and DeepSeek are commonly called open source because their weights are freely available. This is genuinely significant: it means developers can run them without API costs, audit them for safety or bias, customize them for specific tasks, and deploy them privately without data leaving their infrastructure. It also means the community can build on them.

The practical decision for builders: closed API models from OpenAI, Anthropic, or Google are often more capable at the frontier but have cost, privacy, and lock-in tradeoffs. Open-weights models let you self-host (run on your own servers), fine-tune, and avoid per-token fees, at the cost of needing to manage your own infrastructure. Knowing whether a model is open-weights is a key data point when evaluating your stack.