AI model whose trained parameters are publicly released, so anyone can run the model on their own hardware. Not the same as "open source": training data and training code are usually not included.
Definition: AI model whose trained parameters are publicly released, so anyone can run the model on their own hardware. Not the same as "open source": training data and training code are usually not included.
Open-weight means the trained model parameters (the weights, sometimes hundreds of gigabytes of numbers) are publicly downloadable, so anyone can run the model on their own hardware. Major examples in 2026 include the Llama family from Meta and various Mistral and Qwen releases.
For research, open-weight matters in three ways. Privacy: data never leaves your infrastructure, sub-processor lists shrink, and AI Act documentation gets simpler. Reproducibility: you control the model version, no silent vendor swaps. Cost predictability: hardware amortises, while cloud tokens stay per-request forever.
Open-weight is not the same as open source. Open source would release training data, training code, and a licence allowing unrestricted use. Most "open-weight" releases keep training data confidential and place limits on commercial use. Read the model licence before deploying.
This term is referenced in the following articles: