The openness ladder
Models are open to different degrees. This ladder has four tiers, or levels, numbered 0 to 3, plus a 3+ for models trained only on openly licensed data. We checked the examples in October 2026.
| Tier | Meaning | Examples |
|---|---|---|
| 0 Closed | API only | GPT-5, Claude, Gemini, Meta Muse Spark |
| 1 Open weights, custom license | downloadable, use restrictions, not OSI-approved | Llama 3/4 (a cap of 700 million monthly users, naming rules, and an EU exclusion for the multimodal models in the acceptable-use policy), Gemma 1-3, Qwen2.5-72B, Mistral Large 2407, Teuken-7B v0.6 (CC BY-NC) |
| 2 Open weights, OSI license, data and code closed | Qwen3/3.5, DeepSeek V3.2/V4 (MIT), gpt-oss, Phi-4, Mistral Small 3.2, Gemma 4 (Apache-2.0 since April 2026), EuroLLM-9B, Salamandra | |
| 3 Fully open | weights + training code + data or reproducible recipe + checkpoints | Apertus 1.0 and 1.5, OLMo 3, SmolLM3, Marin, Pythia, EuroLLM-22B (claimed) |
| 3+ Open data | trained only on public-domain or openly licensed text | Pleias 1.0 on Common Corpus, Comma v0.1 on the Common Pile (which includes HathiTrust/IA pre-1929 books, Library of Congress digitized books, BHL, Project Gutenberg) |
Apertus reaches tier 3 by naming sources you can obtain yourself and by publishing scripts that rebuild the training data from them. There is no ready-made dataset to download. That is a weaker kind of openness than publishing the dataset itself.
Four yardsticks
Different organizations measure openness in different ways.
- OSI Open Source AI Definition 1.0 (October 2024), from the Open Source Initiative. It requires enough “data information” to rebuild the training set, but not the data itself. Pythia and OLMo pass.
- Model Openness Framework. Class I means weights, code, data and checkpoints (versions of the model saved at points during training) are all released. That is tier 3 on the ladder.
- EU AI Act, Article 53. Since August 2025, every provider of a general-purpose model in the EU must publish a summary of its training content, using the European Commission’s template. This includes providers of open-source models. Fines have applied since August 2026.
- SWIB draft policy. Training data must be available.
Four things that are easy to confuse
- Open weights are not the same as open source. Mixtral, Phi-2 and Llama 2 can all be downloaded, yet none of them meets the OSI definition.
- Fully open is not the same as ethically sourced. Pythia is tier 3, yet its training set, the Pile, contained Books3.
- Open data does not mean a weak model. Comma v0.1, trained only on openly licensed text, matches Llama 2 7B.
- A permissive license is not the same as stable terms. Gemma 3 had custom terms and Gemma 4 is Apache-2.0. Teuken 0.4 was Apache-2.0 and 0.6 is non-commercial. Apertus 1.0 could be downloaded directly, and 1.5 is gated: you must agree to a policy before you can download it.
Check the license of the exact version you download. Terms can change between versions of the same model family.
Exercise: place a model on the ladder
Pick a model you have heard of. Open its page on Hugging Face (the website where most open models are published) and find two things: the license, and whether the training data is listed. Then choose its tier.
Show a solution
Three examples, as of October 2026:
- Llama 3: tier 1. The weights are downloadable, but the license is custom, with use restrictions.
- Qwen3: tier 2. The license is OSI-approved, but the training data and code are closed.
- OLMo 3: tier 3. Weights, code, data and checkpoints are all published.
Check yourself
- I can name the four tiers of the ladder.
- I looked up a license and the training-data statement for one model.
- I can say why “open weights” is not the same as “open source”.
What you should see: one model placed on a tier, with the license and the data statement as your two reasons.