---
title: 'Open-weight model'
description: 'A model whose trained parameters are distributed under stated terms, without implying that its training process or complete system is open source.'
canonical_url: 'https://darkfactory.dev/glossary/open-weight-model'
markdown_url: 'https://darkfactory.dev/glossary/open-weight-model.md'
collection: glossary
date_published: '2026-08-04T00:00:00-04:00'
date_modified: '2026-08-26T00:00:00-04:00'
---

# Open-weight model


## Definition

An open-weight model has trained parameters that recipients can obtain under stated terms. The release may include architecture code, a tokenizer, inference code, checkpoints, or training artifacts, but the label guarantees none of those additions. It also says nothing by itself about commercial use, redistribution, or the right to publish a modified model.

Treat open-weight as an artifact claim followed by a license question: which weights are available, in what format, and what may a recipient do with them? A downloadable file is not enough if the license withholds the intended use.

## Dimensions of an open-weight release

Two open-weight releases can provide very different practical options:

- A release may provide base weights but omit instruction-tuned variants, intermediate checkpoints, optimizer state, or the tokenizer needed to reproduce inputs.
- The license may treat research, commercial use, hosting, modification, and redistribution differently.
- Architecture and inference code may be public while training data, data lineage, filtering, and training code remain undisclosed.
- Local operation still requires compatible serving software, sufficient memory and compute, security maintenance, and evaluation.

Weight access allows direct experimentation, but it does not make the model's learned reasoning legible. Inspecting billions of numerical parameters is different from understanding the examples, incentives, and decisions that produced them. Reproducing the original training run usually requires artifacts far beyond the final checkpoint.

Downloadable artifacts also create a supply-chain obligation. Operators need to verify provenance, serialization safety, runtime dependencies, and update channels rather than assuming weight access makes the package trustworthy.

## Distinguish it from nearby terms

Open-source AI requires broader artifacts and freedoms to use, study, modify, and share. Source-available may expose code or artifacts under restrictions. Proprietary and open-weight are not exact opposites because a vendor can distribute weights while retaining restrictive rights or closed surrounding components. Frontier describes relative capability, not artifact access.

## Check your understanding

A release includes base weights, tokenizer files, and inference code, but no training code or data information. What can an operator now do directly, and which claims about reproducibility or open-source status remain unsupported?

## Also called

open weights model

## Related terms

- [Model weights](https://darkfactory.dev/glossary/weights)
- [Foundation model](https://darkfactory.dev/glossary/foundation-model)
- [Open-source AI](https://darkfactory.dev/glossary/open-source-ai)
- [Proprietary model](https://darkfactory.dev/glossary/proprietary-model)
- [Frontier model](https://darkfactory.dev/glossary/frontier-model)
- [Training data](https://darkfactory.dev/glossary/training-data)

## Related factory areas

- [Model selection, routing & budgets](https://darkfactory.dev/factory/model-routing-budgets)
- [Runtime operations, observability, incidents & healing](https://darkfactory.dev/factory/runtime-operations)
- [Security, privacy & software supply chain](https://darkfactory.dev/factory/security)

## Evidence and further reading

- [Open Source Initiative: Open Weights](https://opensource.org/ai/open-weights)
- [Open Source Initiative: Open Source AI Definition 1.0](https://opensource.org/ai/open-source-ai-definition)
- [Stanford HAI Artificial Intelligence Glossary](https://hai.stanford.edu/ai-definitions)