An open-weights model is an AI model whose learned parameters, the "weights", are published and downloadable, which lets you run it on your own servers and adapt it.
Weights are the output of training: without them, you can only call the model through its publisher's API. With them, you can host it yourself, quantize it so it fits on more modest hardware, or specialize it through fine-tuning. Open weights does not always mean open source: training data and code often remain private, and the license may restrict some uses.
Llama (Meta), some Mistral AI models or Qwen (Alibaba) are well-known examples. These models are mostly distributed through Hugging Face. They appeal to organizations that want to keep their data in-house or control their costs.
Why it matters when hiring
Choosing an open-weights model shifts work to the in-house team: hosting, inference optimization, evaluation, security. You need profiles who know how to deploy and serve a model, not just call an API. The good signal: having run an open model in production with vLLM or an equivalent tool and knowing how to measure its quality. Also check that the candidate reads licenses.