Open weights means an AI model's trained settings are publicly available for anyone to download and run. Open source is stricter: it also requires the training code, data, and methods to be shared. Most AI models marketed as open today are open weights, not fully open source.
How It Works
When an AI model finishes training, what it has learned is stored as numbers called weights — billions of them. Releasing the weights openly means anyone can download the finished model, run it on their own computer, inspect it, and fine-tune it. You get the cake, but not the recipe.
True open source, as defined by decades of software tradition, means sharing the recipe too: the training code, the datasets, the exact process used. With a fully open-source model, you could rebuild it from scratch and verify every claim about it. With open weights only, you can use and study the model, but you cannot reproduce how it was made or audit what data it learned from.
This distinction is not just academic. Regulators, researchers, and businesses care because the recipe reveals risks: was the training data collected fairly? Does it contain private information? Can the safety claims be checked? Open weights give you freedom to use; full open source gives you freedom to verify.
Why It Matters
It matters because the label “open” shapes trust and competition. Open-weights models let startups, researchers, and hobbyists build on state-of-the-art AI without paying a giant lab — a huge democratizing force. Hospitals can run models on their own servers to keep patient data private; developers anywhere can adapt models to their own language.
But the “open” label can also mislead. A company can release weights while keeping its data and methods secret, gaining the goodwill of openness without the accountability. Knowing the difference helps you judge what a release actually gives you — and what it withholds.
Real-World Example
Suppose you are a developer building a customer-support bot for a bank. An open-weights model lets you download it, run it inside the bank’s own data center so no customer data leaves the building, and fine-tune it on banking conversations. That is a big win over sending data to someone else’s cloud. But if a regulator asks, “What data was this model originally trained on?” — with open weights alone, you cannot answer. Only a fully open-source release would let you check.
Popular AI Tools
Notable examples include DeepSeek, whose models are released with open weights; Meta AI, powered by the Llama family of open-weights models; and HuggingChat, an open chat interface for trying many open-weights models in one place.
Frequently Asked Questions
Is Llama open source?
Not strictly. Meta’s Llama models are open weights — you can download and use them — but Meta does not share the full training data and code, so they do not meet the traditional open-source definition. Many people still call them “open” informally.
Why don’t companies release everything?
Training data can contain copyrighted or private material, and the full recipe is a competitive advantage worth a fortune. Some also worry that fully open recipes make misuse easier. So most labs draw the line at releasing weights.
Can I use open-weights models commercially?
Often yes, but check the license — it varies. Some open-weights models allow commercial use freely; others add conditions, such as requiring a separate license once your user base gets large.
Does open weights mean the model is safe?
No. Openness is about access and transparency, not safety. An open-weights model can still produce harmful or biased output, and once weights are public, their safeguards cannot be recalled. Use them thoughtfully.
