Open weights: the model file is only part of the recipe

Open weights means a model's learned parameter files are available under stated terms. They can allow people to run, inspect, adapt or redistribute the model, depending on the license. A hosted API alone does not provide those files.

Weights are the numerical result of training. They do not necessarily include the original training data, filtering rules, optimizer history or enough detail to reproduce that training run.

What you need to run the model

A usable release commonly includes architecture configuration, a tokenizer or processor, weight files and instructions for loading them. These parts must agree: a tokenizer ID refers to the matching model vocabulary, and a weight tensor has a shape expected by the architecture.

A model card can document intended uses, evaluation, training information and limitations. It is useful evidence about a release, but its presence does not establish that every detail of the training process is available.

Availability and freedom are different questions

“Open weights” and “open-source AI” are not interchangeable descriptions. The Open Source Initiative's definition includes requirements beyond parameter availability, including code and information about training data, alongside freedoms to use, study, modify and share.

Read the release's actual terms rather than inferring permissions from a downloadable checkpoint. Restrictions, obligations and available supporting materials differ among releases.

Quantization and LoRA become practical once compatible local parameters and software are available. The distinction matters because running a model, adapting it and reproducing its original training are three different levels of access.