Glossary
Open-weights model
An open-weights model is an AI model whose trained "brain" is published for anyone to download, run, and adapt, like Meta's Llama or Mistral's models. The opposite of closed models like GPT or Claude, which you can only use as a service.
The “weights” are the billions of numbers a model learned during training: the actual brain. Closed models (OpenAI’s GPT, Anthropic’s Claude, Google’s Gemini) keep theirs on their own servers: you rent answers. Open-weights models publish the file: download it, run it on your own machine, fine-tune it, no permission needed.
It’s the difference between a restaurant and a published recipe. The restaurant may cook better, but the recipe is yours: cook at home, change the ingredients, and nobody sees what you’re making.
That last point is the practical headline: an open-weights model running locally sends nothing to anyone, which matters for confidential work. Trade-offs, honestly: the strongest closed models still lead at the high end (the gap keeps narrowing, and Llama, Mistral, DeepSeek, and Qwen are seriously capable), and running big models well takes decent hardware. Note “open weights” isn’t quite “open source”: licenses vary, and training data usually stays private.
Where you’ll meet this
Tools like Ollama and LM Studio that run models on your laptop, Hugging Face (the download hub), and Mistral’s models, Europe’s flagship example. Whether local AI fits your situation is part of what our AI chooser weighs.