An open-weight model is one whose "weights" — the trained numbers that make up the model — are released publicly, so anyone can download it and run it on their own hardware. Meta's Llama, Google's Gemma, Mistral, and DeepSeek are the well-known examples. This is the opposite of a closed model like Claude or GPT, which you can only reach through the provider's API.
"Open-weight" is more precise than "open-source": the weights are shared, but the training data and full recipe usually aren't, so it isn't open source in the traditional software sense.
Why it matters at your desk. For an engineer or a privacy-conscious organisation, open-weight is the answer to a specific question: can the model run somewhere I control? Running it on your own infrastructure means sensitive data never leaves the building — a real consideration for legal, medical, and regulated work — and there's no per-token bill. The trade is that you own the setup, scaling, and inference costs yourself, and the very best frontier models are still typically closed. Releases like Gemma 4 show open-weight models closing the gap.
What to watch for: "open" is a spectrum, not a checkbox. Read the licence — some open-weight models carry usage restrictions (commercial limits, acceptable-use clauses) that matter before you build on them.