Open Weights or Open Source? What the Labels Really Mean

Read the announcements and you would think the AI world has gone gloriously open. Meta, DeepSeek, Alibaba, Nvidia and a wave of Chinese labs all wave the open banner. Most of them are using the wrong word, and the gap between what they say and what they hand over is worth understanding, because it decides what you can actually do with a model.

Two words, two very different things

An open-weight model gives you the finished product. You can download the trained parameters, the billions of numbers that make the model work, run it on your own hardware, and build on top of it. What you do not get is the recipe. The training code, and above all the dataset the model learned from, stay behind the curtain. You have the cake. You cannot see how it was baked, and you cannot bake another.

Open source, in the strict sense, means the recipe too. In late 2024 the Open Source Initiative, the same nonprofit that defined open source for ordinary software, published its Open Source AI Definition. It asks for three things: the model weights under a proper licence, the full training and processing code, and enough detail about the training data that a skilled person could build a substantially equivalent model. Open weights satisfy only the first.

Why almost nothing qualifies

By that standard, most of the famous names are open weight, not open source. Llama, DeepSeek, Qwen and Gemma all let you download and run the model while keeping the data recipe private. Genuinely open-source models that meet the full definition, such as OLMo and Pythia, are rare, and they tend to come from research groups rather than the big labs.

The reason is not mysterious. Training data is where the legal risk and the competitive advantage both live. Publishing the exact corpus invites copyright fights and hands rivals the map to your work. Releasing the weights alone gives developers something genuinely useful while keeping the crown jewels locked away. It is a reasonable business choice. It just is not open source, and calling it that muddies a distinction that matters.

Why the distinction is not pedantry

The label changes what you can trust and what you can reproduce. If you cannot see the training data, you cannot fully audit a model for bias, for memorised private information, or for security holes baked in during training. You are taking the lab's word for what went in. That is fine for many uses and a real problem for others, which is why regulators and security researchers keep pushing on it. Recent releases show the tension clearly, from Meta putting an open-weight agent on ordinary laptops to a near-frontier Chinese model that shipped with almost no safety guardrails.

None of this makes open weights bad. Downloadable models have pushed the field forward and given people an alternative to renting intelligence from a handful of companies. The point is narrower. When a company says open source, it is worth asking which kind, because the honest answer is usually you can have the model, just not the truth about how it was made.