Open-weight Model

A model whose trained parameters are published for download, even if its training data and code aren't.

An open-weight model is one whose developer publishes the trained parameter files publicly, so anyone can download them, run inference locally, and fine-tune them. This differs from open source in the strict sense: the Open Source Initiative's (OSI) definition of open-source AI also requires information about training data and complete training code, whereas many open-weight models only provide the weights and inference code, and their licenses may restrict commercial use or cap the number of users. For embodied-AI researchers, open weights mean an existing VLA can be fine-tuned on their own robot data instead of pretraining from scratch — this is a big part of why the open-source ecosystem has grown so quickly over the past couple of years.

ExamplePhysical Intelligence released π0's weights through openpi, letting researchers download them and fine-tune on their own robot arm data.

Also called

open weights

Related

Open-Source License (Apache 2.0 / MIT / Non-commercial) · openpi (Physical Intelligence) · Pre-training · Fine-tuning · Llama · OpenVLA

Sources

The Open Source AI Definition - Open Source Initiative