What is Open-weight model?

How it works

A trained model is, in the end, a large set of numbers: the weights. With a closed model (most versions of GPT, Claude and Gemini) you can only send requests to the vendor's API. With an open-weight model the vendor publishes the weight files, typically on Hugging Face, with a model card and a licence. You can then:

• run it locally with tools such as Ollama, LM Studio, llama.cpp or vLLM,
• shrink it with quantization to fit consumer GPUs or Macs,
• fine-tune it on your own data,
• host it in your own cloud or data centre.

Open weights vs open source

The terms are often mixed up. The Open Source Initiative's Open Source AI Definition requires the freedom to use, study, modify and share a system, including enough information about the training data and the code used to train it. Many popular open-weight models publish only the weights and inference code, and some use licences with conditions (for example, limits for very large companies or acceptable-use rules). So "open-weight" is the accurate term for most of them; "open source" applies only when the licence and released materials meet that fuller standard.

Why it matters

• Privacy and control: data never leaves your machines, which matters for regulated or client data.
• Cost: no per-token fees once you have the hardware, which pays off at high volume.
• Customisation: you can fine-tune and inspect the model.
• Trade-offs: you handle hosting, updates and safety filters yourself, and the largest open models need serious hardware. The NTIA report also notes that once weights are released they cannot be recalled, which is why policymakers weigh their benefits against misuse risks.

Example

A law firm wants a document assistant but cannot send client files to a third-party API. It downloads an open-weight model, runs a quantized version on an in-house GPU server behind its firewall, and connects it to its document store with RAG.