Getting started in 15 minutes

Two starting points: chat with Meta AI for free, or run a specific Llama model through a hosted API or a local runner. Both paths, step by step.

In this lesson you will

Use Meta AI for a first real task in under five minutes
Set up a hosted-API account and send a first request to a named Llama model
Run a small Llama model locally with a runner such as Ollama

Because Llama is open-weight, "getting started" has two meanings. If you just want to talk to Meta's assistant, you can be done in two minutes. If you want to pick a specific Llama model and control it, you need either a hosted API account or a local runner. This lesson walks through both.

Figure 1.Start free, then add a metered API or a local runner only if you need model choice or privacy. Judge each on the same real task.

Path A: Meta AI, no setup

Chat with Meta AI

1
Open it
Go to meta.ai in a browser, or open the Meta AI chat inside WhatsApp, Instagram or Messenger. Sign in with a Meta account if asked.
2
Do one real task
Paste a paragraph you actually need rewritten, or ask for a plan for something you actually have to do this week. Judge it on that, not on a party trick.
3
Note what you cannot change
There is no model picker and no settings that matter for cost. That is fine for casual use. If you find yourself wanting more control, move to Path B.

Path B: a specific model via hosted API

A hosted API lets you name the exact model, set a system prompt and see a bill. OpenRouter is a convenient host because it serves many Llama models behind one OpenAI-compatible endpoint. Other cloud providers work the same way; the labels below are generic.

Hosted API setup

1
Create an account and add a little credit
Sign up at your chosen host, for example openrouter.ai. Add a small amount of credit, five dollars is plenty, so you cannot accidentally overspend.
2
Create an API key
Find the API keys page in the host's dashboard and create one. Store it as an environment variable, never in code you might share.
3
Pick a model and read its price
Choose Llama 4 Scout or Llama 3.3 70B for a first test. Check the host's own pricing page for that model and compare it with our models list.
4
Send a first request
Use the host's playground if it has one, or the curl example in Lesson 7. Ask the same real task you gave Meta AI and compare.
5
Set a spend limit
Most hosts let you cap monthly spend or set alerts. Do this now, before you forget.

Path C: run it on your own machine

Local runners such as Ollama download a model once and serve it on your laptop. There is no per-token cost and nothing leaves your machine. The trade-off is that you need enough memory: the 8B model runs on many laptops, the 70B model needs a serious GPU or a lot of RAM and will be slow.

bashLocal runner example (Ollama). Model tags and commands are illustrative; check the runner's documentation for current names.

# Install Ollama from its official site, then:
ollama run llama3.3

# For a machine with less memory, start with a smaller model:
ollama run llama3.1:8b

Local setup

1
Install a runner
Install Ollama or a similar runner from its official site. Follow its instructions for your operating system.
2
Pull a model that fits
Start small. If the 8B model runs comfortably, try a bigger one. If your machine struggles, stay small or use a hosted API for the big models.
3
Try the same task
Give it the task you used on Meta AI and on the hosted API. You now have three answers at three different cost profiles to compare.

Settings that matter

Model name: the single biggest cost lever. Check it before every job.
Max output tokens: cap it so a runaway answer cannot cost you a full context window of output.
Temperature: lower for extraction and code, higher for brainstorming. It does not change price, but it changes how often you have to retry.
Spend limits and alerts: set them at the host level, not just in your head.

Knowledge check

You want to try Llama 3.3 70B for free and privately, with no account. Which path fits?

Lesson FAQ

Do I need a Meta account to use Llama?

Only for Meta AI. Hosted APIs need an account with the host, not with Meta. Local runners need no account at all.

Can I run Llama on a normal laptop?

The 8B model runs on many recent laptops. Larger models such as 70B need a lot of memory and are slow without a capable GPU. Use a hosted API for those.

Which host should I use for the Llama API?

OpenRouter is a convenient starting point because it serves many Llama models behind one endpoint. Compare host rates on their own pricing pages and in our [models list](/models).