- Most 'open-source' AI models only release trained weights, not training code, data, or infrastructure details.
- Users can download, run locally, fine-tune, quantize, and build commercial applications with the model weights.
- Apache 2.0 offers maximum permissiveness while Llama licenses restrict high-scale commercial uses and competing AI services.
- Without training data and code, models cannot be independently verified or reproduced by researchers.
- Projects like Pythia and OLMo release complete training materials enabling genuine scientific reproducibility.
The Terminology Confusion
Meta, Mistral, DeepSeek, and others are frequently described as 'open-source' AI labs. This is not quite accurate. In most cases, what's released is the trained model weights — the billions of numerical parametersParametersThe numerical weights inside a neural network that are learned during training — the 'knowledge' of the model, measured in billions for modern LLMs.Learn more → that make the model work. The training code, training data, training infrastructure details, and hyperparameter specifics are usually not released.
True open-source, in the traditional software sense, means code that can be freely used, modified, and distributed — including everything needed to reproduce the work. By that definition, virtually no major AI model is open-source. The Open Source Initiative (OSI) has proposed specific criteria for open-source AI that few current models meet.
What Open-Weight Actually Gives You
With open-weightOpen-WeightA model whose trained weights are publicly available for download, allowing anyone to run, fine-tune, or build on top of it — distinct from fully open-source (which also includes training code and data).Learn more → models, you can: download and run the model locally on your own hardware, fine-tune the model on your own data, quantize for efficiency, build commercial applications (subject to the license), inspect and probe the model's behavior, and share or distribute the weights (again, per license). This is enormously valuable even without full open-source transparency.
Open-weight models have enabled an extraordinary ecosystem: Hugging Face hosts thousands of fine-tuned variants; researchers use them for interpretabilityInterpretabilityResearch into understanding how AI models work internally, including mechanistic interpretability that maps specific behaviors to neural components.Learn more → studies; developers self-host for privacy and cost reasons; companies build specialized vertical AI products. This would not be possible with API-only access.
Why License Terms Matter
Not all open-weight licenses are equal. Apache 2.0 (Mistral Small, Falcon) is maximally permissive — use commercially, modify, distribute, no special requirements. Llama licenses (2 and 4) require specific attribution, prohibit certain high-scale commercial uses, and restrict competing AI service providers. RAIL (Responsible AI License) adds use-based restrictions (no use for weapons, surveillance, etc.).
Before building a commercial product on any open-weight model, read the license carefully. The Llama 4 Community License, for example, explicitly prohibits using the model to power competing AI assistant services. For compliance-sensitive applications, Apache 2.0 models are the safest choice.
Can Anyone Build This?
The fact that training data and code are not released means that even with open weights, no one can independently verify how the model was built or reproduce it. This matters for trust: AI safety researchers cannot fully audit models whose training process is opaque. It also means the 'open' ecosystem depends on the original labs continuing to release weights.
Truly open efforts like Pythia (EleutherAI) and OLMo (Allen AI) release training code, data, logs, and intermediate checkpoints — enabling genuine scientific reproducibility. These smaller models don't compete at the frontier capability level, but they serve the research community's need for transparent, auditable AI.