Llama is Meta’s family of open weight models. Claude, GPT and Gemini are closed source: you send a query, it goes to their cloud, an answer comes back. Llama gets downloaded onto your machine and runs there.
Open weight vs open source: what’s the difference?
Open weight means you get the model file. The weights, which are the trained numbers that make the model do what it does, sit on your hardware. You run the model yourself, you can tune those weights, and no API tokens get billed for any of it.
Open source would mean more. It would mean the training data as well, and a licence with no restrictions on what you do with any of it. Llama gives you neither. You get the finished model. You do not get the recipe. Meta’s bet gets described as a bet on open source, but the thing you can download is the weights.
What you give up against a closed model is capability. Llama models are significantly smaller than the closed frontier models, so they are less intelligent, slower and less capable. They do not compete on quality. They compete on safety and on cost.
Is Llama free?
Yes. Free to download, free to run, and free for commercial use since Llama 2. Nothing is billed per token.
The cost moves rather than disappears. Instead of paying a provider by the token, you pay for hardware: a machine with enough RAM to hold the model, or RAM rented from a cloud provider. Run a thousand requests this month or a million, and as long as the RAM copes, the invoice is the same.
How Llama got to version 4
Meta has been at this longer than the release history suggests. Its research department, FAIR, started in 2013. In 2025 Meta restructured around a new group called the Meta Superintelligence Lab, led by Alexandr Wang, the former CEO of Scale AI, one of the biggest data labellers. They have been quiet since, which is why there is so much anticipation about what comes next.
Public releases only began in 2023.
| Model | When | What it is good for |
|---|---|---|
| Llama 1 | 2023 | Researchers only. It leaked, so everyone else got it anyway. |
| Llama 2 | Late 2023 | The first official open weight release, free even for commercial use. Fine for a small local job, struggles with complex workflows. |
| Llama 3 | 2024 | The biggest jump. Reasoning capability and light coding work, closing most of the gap to the frontier. |
| Llama 4 | 2025 | Roughly a GPT-4 competitor. Still about a year behind the frontier, and free. |
A few months from Llama 1 to Llama 2, then a year to Llama 3, then another year to Llama 4.
What hardware do you need to run Llama locally?
Not the laptop you already own. Llama 4 is cumbersome on a normal computer. You need a beefy machine with a lot of RAM, and that is what keeps local Llama out of daily use for most people.
The workaround is to rent RAM from a cloud provider like AWS and run the model on your own cloud infrastructure instead of on the machine under your desk. You still control the deployment and the data path. You just do not have to buy the hardware.
The trend runs your way. Each generation gets stronger while the RAM it needs drops. In a couple of years you could be running something with the capability of today’s Sonnet 4 or Sonnet 5 on your own machine.
That is Meta’s long bet, and it is how they intend to put the others out of business. You pay API prices only for frontier work. The agent that moves a job in your calendar or drafts your emails runs locally for nothing.
What Llama is good at
Self-hosting. You run the servers and the data processing yourself. In an organisation that means the racks in the back office. The data never leaves the network, which is what makes it usable in high-risk industries like finance and healthcare.
Cost predictability. You know what those servers cost you. A thousand requests or a million requests, the bill is the same as long as the RAM can handle the load. The job then is getting more work out of a fixed monthly cost.
Customising. Because the weights are open, you can change them to influence how the model responds. If you want more data-backed answers, you tune for that rather than front-loading a pile of instructions into every prompt.
Set against a closed frontier API, the trade looks like this.
| Self-hosted Llama | Closed frontier API | |
|---|---|---|
| Quality | About a year behind the frontier | The frontier |
| Where your data goes | Never leaves your network | Their cloud |
| Cost shape | Fixed monthly server cost, the same for a thousand requests or a million | Billed per token |
| Setup and upkeep | Cumbersome, and you maintain it | Nothing to run |
Should you self-host an AI model?
For a solar or HVAC operator the question is which jobs justify the hardware. Drafting customer follow-ups and updating the crew calendar are local-model jobs, and a Llama box in the back office does them at a cost you already know. Quoting and anything a customer judges you on goes to a frontier model, and you pay per token for it.
Most operators will land on that mix rather than picking a side. Apple is making the same bet from the other side: it has released no models of its own and is betting instead on selling the hardware that models run on. But setting up a local stack and maintaining it is cumbersome, and that alone sends plenty of people back to a cloud provider.
Meta is not the only source of open weights either. The Chinese labs release open weight models too, so there is real competition for the same position.
