Definition
A model whose trained parameters are available to download, inspect, or run under stated terms. The release may also include architecture code, inference code, documentation, or training artifacts, but those additions are not guaranteed by the label.
Open-weight is best treated as a factual statement about artifact availability followed by a license question: which weights are available, in what format, and what may a recipient do with them?
Dimensions of an open-weight release
Two releases both described as open-weight can provide very different practical freedoms:
- Artifact completeness: base weights, instruction-tuned weights, checkpoints, tokenizer, configuration, and optimizer state may be released selectively.
- License permissions: use, commercial use, modification, fine-tuning, redistribution, and hosting may carry different conditions.
- Training transparency: architecture and inference code may be available while training data, data lineage, filtering, and training code remain undisclosed.
- Reproducibility: possessing weights permits inference but rarely permits reproducing the original training run.
- Operability: local execution still requires compatible serving software, sufficient compute and memory, security maintenance, and evaluation.
- Supply-chain trust: a downloadable artifact can be inspected more directly, but operators also inherit responsibility for provenance, serialization safety, dependencies, and updates.
Weight access creates options; it does not establish that every recipient can practically exercise them.
Distinguish it from nearby terms
- Open-source AI requires broader artifacts and freedoms to use, study, modify, and share. An open-weight release does not automatically meet that threshold.
- Source-available may expose code or artifacts under restrictions. Open-weight identifies the weights specifically.
- Proprietary model and open-weight are not perfect opposites. A vendor can release weights under materially restrictive terms or retain proprietary surrounding components.
- Frontier model describes relative capability, not artifact access or licensing. A model can be both frontier and open-weight.
Check your understanding
Read the license and inventory the released artifacts before assuming what "open" permits, reveals, or makes reproducible. If only the checkpoint is available, say open-weight, not open-source, unless the broader requirements have also been established.