Llama

Llama is a family of open-weight foundation models created by Meta, downloaded over one billion times with more than 85,000 derivative models published on Hugging Face. The latest generation, Llama 4, uses Mixture-of-Experts architecture with early fusion to jointly pre-train on unlabeled text and vision data. Llama 4 Maverick has 17 billion active parameters across 128 experts and scores 80.5 on MMLU Pro, 69.8 on GPQA Diamond, 43.4 on LiveCodeBench, 73.4 on MMMU, 73.7 on MathVista, 90.0 on ChartQA, and 94.4 on DocVQA, with a 10-million-token context window. Llama 4 Scout fits on a single Nvidia H100 GPU and scores 74.3 on MMLU Pro and 57.2 on GPQA Diamond. The Llama 3 generation spans 1-billion to 405-billion parameters, where the 70-billion Llama 3.3 matches the 405-billion Llama 3.1 on knowledge benchmarks while costing one-sixth as much to run. Enterprise case studies show measurable returns. Shopify increased token throughput by 76 percent and achieved 97.7 percent Macro-F1 on intent detection while cutting compute costs by 33 percent. Stoque reduced repetitive customer support queries by 50 percent and increased task completion by 30 percent, yielding an 11-point satisfaction improvement. Oxide AI automated internal workflows to reduce manual processing by 60 percent. Estimated inference cost is 0.19 to 0.49 dollars per million tokens for distributed deployment. Models are available on Hugging Face, Azure AI, AWS Bedrock, Google Cloud Vertex AI, NVIDIA NIM, Groq, Together AI, and through the Llama API. Meta provides Llama Guard for input-output safety filtering, Prompt Guard for injection defense, and the Llama Evaluation harness for benchmark replication. The Acceptable Use Policy permits commercial use with over 700 million monthly active users requiring a license.