AI model families: Meta's Llama (open source)

Llama is Meta's AI model and you can download the actual model weights. That means hobbyists, researchers, and small companies can run real AI without paying OpenAI. It's a big deal for the field.

Some examples

• Run Llama on your laptop with Ollama
• Use Llama-based tools that don't ship your data anywhere
• Build a Discord bot that runs locally
• Compare Llama 3 vs Llama 4 quality

Try it!

Install Ollama. Download a small Llama model. Ask it a question — entirely on your laptop, no internet needed for the answer. Wild, right?

Llama is Meta's family of open-source models. The weights are public, so devs can download them, fine-tune them, and run them on their own computers. That means total privacy and no API bills.

Some examples

• Run Llama 3 locally on a decent laptop using Ollama (free).
• Use a fine-tuned Llama for niche stuff (medical, legal, etc.).
• Llama 70B competes with the smaller paid models on quality.
• Open weights means devs can audit the model — closed ones can't.

Try it!

Install Ollama (free) and download a small Llama model. Chat with it offline. Notice it works with no internet.

Llama 3.3 from Meta is open-weight — you can download it and run it yourself. The 70B model on a beefy Mac matches GPT-4-class quality for many tasks. Privacy and zero per-token cost are huge.

Some examples

• Run Llama 3.3 locally with Ollama or LM Studio.
• Use Groq or Together AI for blazing-fast hosted Llama.
• Open weights means you can fine-tune it on your own data.
• Llama works offline — perfect for sensitive personal data.

Try it!

Install Ollama and run Llama 3.3 8B on your laptop. Ask it a question. Notice it works with no internet.

Meta releases Llama models with the weights public — meaning anyone can download them, run them locally (via Ollama), fine-tune them, or build companies on them. Even if you never use Llama directly, you benefit: when a free, open option exists, paid models have to keep prices down and quality up. Llama is the industry's pressure release valve.

Some examples

• DeepSeek, Qwen, Mistral, and others are also open-weight — Llama isn't alone, but it's the most famous.
• Many startups (Perplexity, Groq) run Llama as a cheap backbone for their products.
• Researchers use Llama because they can see and study every weight — closed models hide that.
• You can fine-tune Llama on your own writing for a personal AI — impossible with GPT-5 or Claude.

Try it!

Read the latest Llama release post on ai.meta.com. Notice how they describe the trade-offs vs closed models.

Closed models (Claude, GPT) live on company servers. Open-source models (Llama, DeepSeek-V3, Qwen) can be downloaded, run on your own hardware, fine-tuned, and used offline. Quality has gotten close to closed models, especially for coding (DeepSeek) and Chinese tasks (Qwen).

Some examples

• You run Llama 3 on your laptop with Ollama for offline use.
• You fine-tune DeepSeek for your company's domain — closed models can't do that.
• You use Qwen because it's the strongest open model for Chinese.
• You use Llama because it's free and your project can't send data to OpenAI.

Try it!

Install Ollama and run Llama 3 (or any small model) on your own machine. Ask it a question. Notice the speed and that it's all local.

Open-source models like Llama, Mistral, and Qwen don't beat the frontier closed models on raw capability — but they win on three things: data privacy (nothing leaves your machine), cost at scale, and the ability to fine-tune for your domain.

Some examples

• You run Llama 3 locally via Ollama for sensitive notes that can't go to a cloud API.
• Mistral 7B handles a high-volume classification task at $0 per call after the GPU cost.
• A fine-tuned Qwen on legal docs outperforms GPT-4 on your specific domain.
• Llama running on your laptop gives offline coding help on a flight.

Try it!

Install Ollama and pull a Llama or Mistral model. Run a prompt locally. Note the speed and quality.

local models are slower but private and free

Some examples

• Using Ollama to install in one command
• Picking a 3B model for a 16GB laptop
• Knowing it will not match Claude on hard stuff

Try it!

Open your favorite AI tool and try one of the examples above. Pick the one that matches what you are actually working on this week. Spend 10 minutes, no more. Notice what worked and what did not — that's the real lesson.