☀️ AI Morning Minute: Open Weights
“Open source” and “open weights” sound like the same thing. They’re not, and the difference matters more than you’d think.
When an AI lab releases a model to the public, the question isn’t just whether it’s free. It’s what exactly they’re releasing. Open weights means the trained numerical parameters of the model are publicly available, so anyone can download and run it.
That’s different from open source, which would include the training code, the data, and the recipe. Most models people call “open” are actually open weights. The distinction shapes what you can do with them, what you can trust about them, and who controls them.
What it means
Every AI model is built from billions of numerical parameters called weights. These are the values the model learned during training, and they’re what make it capable of generating text, writing code, or answering questions. A closed model keeps those weights locked up. You access it through an API, send your data to someone else’s server, and get a response back. You have no visibility into how it works and no ability to run it yourself.
An open-weights model releases those parameters publicly. You can download them, run the model on your own hardware, fine-tune it on your own data, and modify its behavior. Meta’s Llama, Mistral, DeepSeek, and Google’s Gemma are all open-weights models. What they don’t release is the training data or the full training code, which is why they’re not technically open source by the strict software definition.
Why it matters
The gap between open-weights and closed models is closing fast. As of mid-2026, five of the top ten models on major AI benchmarks are open-weights. A year ago, the frontier was exclusively closed. The competitive pressure from open-weights releases has driven down API prices and forced closed labs to accelerate their own releases. DeepSeek’s open-weights release in early 2025 alone wiped roughly $600 billion from NVIDIA’s market cap in a single day.
For regulated industries, open weights aren’t optional. A hospital, a bank, or a government agency can’t send sensitive data to an external API. Open-weights models let them run AI entirely on their own infrastructure, with no data leaving their systems. The EU AI Act now requires that open-weights GPAI models make their parameters, architecture, and usage information publicly available, which is turning auditability from a nice-to-have into a compliance requirement.
Sovereignty is the real driver. Countries and enterprises that depend entirely on closed models from foreign companies are exposed to pricing changes, access restrictions, and geopolitical risk. Open weights let you own the model. That’s a different kind of control than renting access through an API.
Simple example
A closed model is like streaming music. The songs exist, you can listen, but you don’t own the files and the service can change the terms whenever it wants.
An open-weights model is like buying the album.
You have the files. You can play them anywhere, remix them, share them, or modify them. You still didn’t write the songs, and you don’t know exactly how they were recorded. But the recording is yours.