Menu
Concepts

Open-weight models

3 min readbeginnerUpdated 28 Sept 2026
1 · In one line

An open-weight model makes its trained weights publicly available for download.

1 · What it is

Open-weight means the model’s weights are publicly available for download. The recipient can download and run the model locally.

The label does not settle every openness question. The International AI Safety Report says open-weight models can be, but are not necessarily, open source. It also says that the full benefits of AI openness require more than sharing weights alone.

Licenses matter too. Mistral 7B was released under Apache 2.0. OpenAI describes its gpt-oss weights as downloadable under Apache 2.0 and a usage policy. Jukebox released both model weights and code.

2 · Why it exists

Weight access changes who can run a model locally.

DownloadOpen-weight users can download and run the model locally.
LicenseWeight availability and permission to use or modify are separate questions.
ReproduceThe International AI Safety Report says full AI openness requires more than sharing weights alone.
3 · How it works

Separate the released artifact from the permissions and supporting materials.

Open-weight means weights are downloadable; it does not by itself establish a fully open-source release.
  1. 1 · obtainDownload the published model weights.
  2. 2 · inspectRead the license and usage terms attached to the release.
  3. 3 · runExecute the model on infrastructure you control.

The International AI Safety Report says open-weight models can be, but are not necessarily, open source.

4 · Where it's used
WhoWhat they askWhat it works with
Application developer“Can I run this model outside the publisher's service?”Downloadable weights and inference requirements
Legal team“What uses and redistribution does the license permit?”The model license and usage policy
Researcher“Are training data and code available too?”The complete release artifact list
5 · What it solves, and what it doesn't
solves
  • Downloadable weights let recipients run a model locally.
  • A release can pair weights with a permissive license.
  • OSI identifies trained parameters as the preferred form for modifying machine-learning weights.
  • Jukebox provides an example of releasing model weights and code together.
doesn't solve
  • Open-weight does not necessarily mean open source.
  • Full AI openness requires more than sharing model weights alone.
  • A license can still place conditions on use.
6 · Go deeper
WeightsParametersBase modelsInferenceModel serving

Sources used

This explainer is written in original language. The links below support its factual claims.

  1. officialInternational AI Safety Report 2025, UK Government · read 28 Sept 2026
  2. officialOpenAI open-weight models, OpenAI · read 28 Sept 2026
  3. officialOpen Source AI Definition 1.0, Open Source Initiative · read 28 Sept 2026
  4. paperMistral 7B, Jiang and colleagues · read 28 Sept 2026
  5. paperJukebox: A Generative Model for Music, OpenAI · read 28 Sept 2026
Suggest a correction