The open model from Meta, the company behind Facebook and Instagram. You download it once and it's yours forever, no subscription, no stranger's server in the middle.
Technically, Llama is what's called open weight. The weights are the thousands of billions of numbers the model learned during its training, the core of its intelligence. Meta publishes them freely to download, along with instructions on how to run them. You don't pay a subscription, you don't go through someone else's server, you're not at the mercy of whether the service shuts down tomorrow. Download the files once and the model is yours forever, even offline.
In practice this means a small company can build a chatbot or document analysis tool powered by Llama, installed on its own server, without paying anyone per request. A hospital that doesn't dare send patient data to someone else's cloud can run Llama behind its own firewall. A developer can modify it, retune it for a specific task, embed it in their own product. None of this is possible with ChatGPT, which remains a closed product of OpenAI.
Why Meta does this for free is pure business calculation. The company fell behind in the race for its own mass-use AI product, but it has something else: money, servers and a wish for the whole world to build on its foundation. If thousands of companies and researchers get used to Llama, Meta becomes the standard the whole ecosystem revolves around, instead of watching from the sidelines while OpenAI and Google dictate the rules. On top of that, the model powers its own products - the AI features in WhatsApp and Instagram come out cheaper when the foundation is yours, not rented.
There's a catch too. The license is only free up to a point - a company with more than 700 million monthly active users has to make a separate deal with Meta to use Llama. In other words, the blueprint is free for everyone, except those who would themselves become a real competitor to Meta.
Let's skip the illusions
It's expensive and risky to be tied to someone else's closed model that might raise prices or cut off access tomorrow. That's why open-weight models like Llama interest me directly, not as news, but as a tool. The difference between an open and a closed model isn't a technical detail for engineers. It decides who controls the technology tomorrow - only the big platforms, or also the small player with a good idea.
So I look at Llama soberly. It's useful, it's real, but it isn't an act of generosity. Meta is playing the long game, giving out the blueprint today so it's the one who decides the next floor of the building.