ML//open-weight model

An open-weight model is a machine learning model whose trained parameter files are distributed under specified license terms, which enables local inference, fine-tuning, quantization, inspection and deployment without sending every request to the original provider. A closed model keeps the weights and exposes its behaviour only through an interface the provider controls; holding the weights is also what permits quantization, pruning and the other deployment-specific transformations of model compression.

An open-weight model is a machine learning model whose trained parameter files are distributed under specified license terms, which enables local inference, fine-tuning, quantization, inspection and deployment without sending every request to the original provider. A closed model keeps the weights and exposes its behaviour only through an interface the provider controls; holding the weights is also what permits quantization, pruning and the other deployment-specific transformations of model compression.

Open-weight does not automatically mean open source. Training data, training code, architecture details or commercial rights may remain restricted. Nor does weight access remove dependency on accelerator vendors, serving frameworks and upstream model releases: local deployment reduces the dependency on an API and leaves the other layers of vendor lock-in in place.

Possessing the engine gives important freedom. It does not mean the entire factory came with blueprints.