Language models & chatbots
An open-weights model is one whose trained parameters have been published, so anyone can download it, run it on their own hardware, and modify it.
Most flagship AI models are closed: you can use them only through the company's API, and the numbers inside are secret. An open-weights model is the opposite. The creator releases the weight file, often tens of gigabytes, and anyone can run it locally, fine-tune it, or build a product on it without sending data to a third party.
This is not quite the same as open source. Open source usually means the training data, code, and full recipe are shared too. Many open-weights releases share only the weights and attach a license with conditions. Still, they let students, researchers, and companies experiment freely.
As of 2026, notable open-weights families include Meta's Llama, Mistral, DeepSeek, Alibaba's Qwen, Google's Gemma, and OpenAI's gpt-oss. Smaller versions run on a laptop or phone; the largest need serious GPU hardware. Open weights also raise safety debates, since safeguards can be removed by whoever holds the file.
A student downloads a small open-weights model, runs it on a gaming laptop with no internet connection, and builds a private study assistant.