Difference between open weights and open-source AI

Open weights and open source describe how much of an AI model its maker shares. With open weights you can download the trained model and run it yourself, but you cannot see the data and code that made it. Open-source AI shares those as well, so others can check and rebuild the model. To run a model in-house, open weights are usually enough; the licence decides what you may do.

What does open source require?

The weights are the numbers a model learned during training; with them and the code that runs them, you can use the model or fine-tune it. The Open Source Initiative (OSI) sets a higher bar in its Open Source AI Definition 1.0 of 28 October 2024. Anyone must be free to use the system for any purpose without asking permission, and to study, change and share it. The release must include the complete code to train and run it, and a description of the training data detailed enough for a skilled person to build a substantially equivalent system. Data that cannot be shared must still be described. Weights on their own, which the OSI calls open weights, lack exactly those parts.

The licence counts as much as the parts. In July 2024 Meta called Llama 3.1 405B "the first frontier-level open source AI model". In February 2025 the OSI wrote that the Llama licences are not open source, because they discriminate against users and restrict what the model may be used for. Companies that had more than 700 million monthly active users when Llama 3 came out need a separate licence from Meta (Llama 3 licence). For Llama 3.2's multimodal models, the rights are not granted to people and companies based in the EU; end users of a product built on them are exempt (use policy).

Which models meet the definition?

While drafting the definition, the OSI tested it on a number of models. Pythia (EleutherAI), OLMo (Ai2), Amber and CrystalCoder (LLM360) and T5 (Google) passed. BLOOM (BigScience), StarCoder2 (BigCode) and Falcon (TII) would probably pass with different licence terms. Llama 2 (Meta), Grok (xAI), Phi-2 (Microsoft) and Mixtral (Mistral) did not. The OSI stresses that these results are not certifications (FAQ).

Ai2's Olmo 3 (November 2025) shows what a fully open release contains: the weights and intermediate checkpoints, the training data, and the code for training and evaluation. The weights and code are under the Apache 2.0 licence, the pretraining data Dolma 3 under the ODC-BY data licence. Many popular models are open weights under a permissive licence, such as DeepSeek-R1 (MIT) and Gemma 4 and OpenAI's gpt-oss (Apache 2.0): you may use them commercially, but their training data is not public. Anthropic, the maker of Claude, publishes no open-weight models (September 2026).

Are there levels in between?

The Model Openness Framework (White et al., Linux Foundation, 2024) ranks releases in three classes. Class III, open model, covers the architecture, the final weights and documentation such as a technical report, evaluation results, model cards and data cards. Class II, open tooling, adds the code for training, evaluation and inference. Class I, open science, adds the training datasets, a research paper and intermediate checkpoints. A release under a licence with use restrictions does not reach class III.

Example

Suppose a company has to show an auditor where its AI model's knowledge comes from. For Llama 3 it has only Meta's description to go on: more than 15 trillion tokens from publicly available sources (model card). With Olmo 3 it can download the training data itself, and Ai2's OLMoTrace highlights passages in an answer that appear word for word in that data.