Llama: Meta's Open-Source LLM Explained

Llama (Large Language Model Meta AI) is Meta's family of open-source large language models. Since its release in 2023, Llama has become the foundation of the open-source LLM ecosystem — more fine-tunes, tools, and integrations are built on Llama than any other model family.

The Llama timeline

• Llama 1 (Feb 2023) — 7B to 65B parameters, leaked then open-sourced. Sparked the open-source LLM movement.
• Llama 2 (Jul 2023) — 7B, 13B, 70B. First version with commercial use license. Included chat-tuned variants.
• Llama 3 (Apr 2024) — 8B and 70B. Dramatic quality improvement. Competitive with GPT-3.5 at 8B size.
• Llama 3.1 (Jul 2024) — Added 405B model. 128K context window. First open-source model to challenge frontier models.
• Llama 3.2 (Sep 2024) — Multimodal (11B, 90B vision models). Lightweight 1B and 3B for edge deployment.
• Llama 3.3 (Dec 2024) — 70B model matching Llama 3.1 405B quality at lower cost.

Llama 3.3 in 2026

Llama 3.3-70B is the current go-to open-source model for most production teams. It scores 86% on MMLU, handles 128K context, and runs on 40GB VRAM (two RTX 4090s or a Mac Studio M2 Ultra).

Running Llama locally

Ollama is the easiest way to run Llama models locally — see our Ollama models guide for the full setup. For production serving at scale, vLLM is the standard choice.

ollama pull llama3.2        # 3B — runs on any laptop
ollama pull llama3.2:8b     # 8B — needs 8GB RAM
ollama pull llama3.3:70b    # 70B — needs 40GB VRAM
ollama run llama3.2

Llama via API

You don't have to self-host. Llama is available via API through several providers:

• Groq — fastest inference (~250 tok/s for 70B). See our Groq guide.
• Together AI — competitive pricing for Llama models
• Replicate — pay-per-second GPU billing
• Meta AI — direct access via meta.ai

Llama vs GPT-4o vs Claude