Llama

Overview

This is a model family overview. For version-specific details, see the individual model entries linked below.

Llama is Meta’s open-weight model family and the most widely adopted open-source model series. It has driven much of the local and self-hosted AI ecosystem, and the Llama 4 generation shifts the family more explicitly into natively multimodal MoE models rather than only text-first releases.

Current Latest

Llama 4 is the current generation, led by Llama 4 Scout and Llama 4 Maverick. Meta’s official April 5, 2025 launch materials describe both as natively multimodal open-weight models with 17B active parameters, using 16 experts for Scout and 128 experts for Maverick.

Strengths

• Fully open weights — can be downloaded, hosted, and fine-tuned under Meta’s license terms
• Excellent option for local and on-premises deployment
• Strong privacy properties — data never leaves your infrastructure
• Native multimodal support in the Llama 4 generation
• Vibrant community ecosystem across quantized releases and local serving stacks

When to Choose Llama

• Privacy-critical workloads — healthcare, legal, finance where data must stay local
• Self-hosted deployments — on-premises or private cloud setups
• Fine-tuning — custom models for specific domains or tasks
• Cost optimization — no per-token API costs after infrastructure investment
• Offline/air-gapped environments — works without internet access

Running Locally

The easiest ways to run Llama locally:

• Ollama — Single-command setup, CLI-based, great for development
• LM Studio — Desktop app with GUI, good for exploration

Access

• Download from Meta (Hugging Face, direct)
• Via Ollama, LM Studio, and other local inference tools
• Cloud providers: AWS, Azure, Google Cloud, Together AI, Fireworks AI
• API access through many third-party providers