The Openness Ladder: 4 levels of open-source AI
The phrase "open source AI" now covers at least four different things. A model you download and modify is not the same as a model you download but may not resell. If you cannot place a model on a ladder, you cannot make a safe decision about it. So we built one.
Four meanings, one word
When a vendor says "our model is open", they might mean any of these: the weights are published but the license is custom; the weights are under a permissive license; the training code is public; or the whole thing, data included, is reproducible. Each meaning changes what you are allowed to do and what you are relying on.
The Openness Ladder
We grade every model we write about on a six-rung ladder, from closed to fully open. The full rubric lives on our framework page. The short version:
- L0 Closed API. Nothing is published. You rent access and own nothing.
- L1 Open weights, restricted. Weights are downloadable, but a custom license caps use. Example: Llama 4, Gemma 3.
- L2 Open weights, permissive. Weights under Apache 2.0 or MIT. Commercial use is allowed. Example: Qwen3, DeepSeek.
- L3 Open code. Training and inference code are public, not just the weights.
- L4 Open data. The training data is published or fully documented. Example: OLMo 2.
- L5 Fully open. Weights, code, data and license are open end to end. Example: OLMo 2, BLOOM, Pythia.
What 2025 taught us
The 2025 release wave made the ladder necessary. Meta’s Llama 4 ships under a Community License with a monthly-active-user threshold above which you must ask for terms. Google’s Gemma 3 ships under the Gemma Terms, which restrict use on larger services. By contrast, Alibaba’s Qwen3 and DeepSeek’s V3 and R1 publish under Apache 2.0 and MIT, which let you build and sell freely. Allen AI’s OLMo 2 goes further and publishes the data. Same word, four different legal positions.
The Open Source Initiative has been clear that "open source" should mean an OSI-approved license. Several 2025 models fail that test while still being called open. The ladder is how we keep the record straight without arguing about definitions.
How to use it
Before you adopt a model, grade it. If your use is a small internal tool, an L1 or L2 model may be fine. If you are building a product you intend to sell at scale, you need to read the license threshold, because L1 models can force a commercial agreement past a usage line. If you are a public body that must show reproducibility, you should be looking at L4 or L5. Grade first, build second.
The Openness Ladder is the grading step of a loop that re-examines one research base for private, enterprise and non-profit buyers. It adapts the producer-side Model Openness Framework into a buyer-side grade you can use before procurement.