Models
Weights
Also known as: model weights
The learned numbers inside a neural network that store what it has learned; "open weights" means a model whose weights are publicly available.
The weights are the numbers inside an AI model that hold everything it learned during training, effectively the model's "brain state". When people talk about "open-weight" models, they mean the company published these numbers so anyone can download and run the model themselves, as opposed to "closed" models you can only use through a company's service.
Weights are the trainable values in a neural network, the strengths of the connections, that training adjusts and that collectively encode what the model has learned. They are essentially the same thing as the model's parameters.
Open vs closed weights
- Open-weight models: the weights are published (e.g. Llama, Mistral). You can download, run, fine-tune and self-host them, greater control, privacy and customisation.
- Closed models: weights are kept private; you access the model only via the provider's API.
This distinction shapes cost, control, privacy and the open-source AI ecosystem. Note "open weights" is not always the same as fully "open source" (which would also include training data and code).
Weights are the learned entries of a network's weight matrices (and biases), synonymous in practice with its trainable parameters; they constitute the model's parametric memory produced by training. The open- vs closed-weights axis is a defining structural feature of the ecosystem: released weights (e.g. Llama 2; Touvron et al. 2023, Mistral) enable self-hosting, fine-tuning/LoRA, quantization, on-prem privacy, and research access, whereas closed models are consumable only via API, trading control for managed convenience and (often) frontier capability.
Terminology precision matters: "open weights" denotes downloadable parameters under some licence, which may restrict use, and differs from fully "open source" (weights plus data, code and full reproducibility). Considerations include licence terms (commercial use, output restrictions), provenance and safety (released weights cannot be recalled, complicating misuse mitigation), and the strategic/geopolitical dimension of who controls frontier weights. Open weights underpin much downstream innovation, customisation and cost-efficiency.