Meta Llama: Everything you need to know about the open generative AI model (techcrunch.com)

Meta’s Llama is an open-weight family of generative AI models — most recently Llama 4 (April 2025) — that Meta distributes for download and cloud hosting via partners like AWS, Google Cloud, Microsoft Azure and Hugging Face. Llama 4 comes in three announced variants: Scout (17B active / 109B total, 10M-token context, 16 MoE experts) for huge-document workflows; Maverick (17B active / 400B total, 1M-token context, 128 experts) as a generalist for coding and assistants; and the unreleased Behemoth (288B active / 2T total, 16 experts) positioned as a teacher model. All Llama 4 models are natively multimodal (text, image, video), trained on large unlabeled multimodal corpora and 200 languages, and support third‑party tool plug‑ins (Brave Search, Wolfram Alpha, Python interpreter) when configured. Why it matters: Llama’s openness gives researchers and developers control for fine-tuning, on‑prem/cloud deployment, and model distillation—accelerating experimentation outside API‑only ecosystems. Meta also supplies safety and security toolkits (Llama Guard, Prompt Guard, CyberSecEval, Llama Firewall, Code Shield) and a cookbook for adaptation. Important caveats: Meta’s training data includes scraped/pirated texts and social-media content, raising copyright and privacy concerns; license terms require special approval for apps with >700M monthly users; and Llama’s code-generation and hallucination performance lag top competitors (e.g., Maverick scored 40% on LiveCodeBench vs GPT‑5’s ~85%). Users should validate outputs and apply Meta’s guardrails when deploying production systems.