AI Model Wiki

    Skip to main content
    6 min read
    Lesson 1 of 6

    Introduction to Meta Llama

    From the course Meta Llama for Developers

    Meta's Llama (Large Language Model Meta AI) is one of the most important developments in AI — a family of powerful, open-weight language models that anyone can download, run, and modify.

    Why Llama Matters

    Before Llama, the most capable AI models were locked behind proprietary APIs from OpenAI, Google, and Anthropic. Llama changed the game by making high-quality models freely available:

    • Open weights — Download and run the full model on your own hardware
    • Commercial license — Use in production applications (with some conditions)
    • Community ecosystem — Thousands of fine-tuned variants on Hugging Face
    • No API costs — Run locally with zero per-token charges

    The Llama Family

    • Llama 2 (2023) — The model that kicked off the open-source AI revolution. Available in 7B, 13B, and 70B parameter sizes.
    • Llama 3 (2024) — Major quality improvement with 8B and 70B sizes. Competitive with GPT-3.5 on many benchmarks.
    • Llama 3.1 (2024) — Introduced the massive 405B parameter model, rivaling GPT-4. Extended context to 128K tokens.
    • Llama 4 (2025) — Latest generation with Mixture-of-Experts architecture, multimodal capabilities, and up to 2M token context.

    Model Sizes Explained

    The "B" stands for billions of parameters:

    • 8B — Runs on a good laptop/desktop GPU (16GB+ VRAM). Fast, good for many tasks.
    • 70B — Needs high-end GPU(s) or cloud. Excellent quality, approaching frontier models.
    • 405B — Requires multi-GPU setups. Top-tier quality rivaling proprietary models.

    Choose based on your hardware and quality requirements.

    Why Open Weights Change the Game

    Proprietary APIs mean your application can change overnight if the provider changes pricing, deprecates a model, or goes down. With Llama, you download the weights once and they are yours forever. You can serve them on your own infrastructure, fine-tune them on your data, and distribute them as part of your product. This permanence is why companies like Spotify, Goldman Sachs, and Shopify have adopted Llama for internal tooling where they cannot accept dependency on a third-party API.

    Common Pitfall: Instruction-Tuned vs. Base Models

    Llama is released in two variants: base (pre-trained) and instruct (fine-tuned to follow instructions). New developers frequently download the base model and are surprised when it responds with raw token continuation rather than answering questions. Always use the "-Instruct" variant for conversational and task-following applications. Use base models only if you are fine-tuning from scratch on your own task data.

    Key Takeaways

    • •Llama is Meta's open-weight LLM family — free to download and use commercially
    • •Models range from 8B (laptop-friendly) to 405B (frontier-quality)
    • •Llama 4 brings Mixture-of-Experts and multimodal capabilities
    • •Open weights enable local deployment, fine-tuning, and zero API costs
    0 of 6 lessons marked complete

    Progress is kept in this browser only.