Llama 4
summary
Llama 4 launches two models, Scout and Maverick, that use mixture‑of‑experts with about 17 billion active parameters (Scout uses 16 experts, Maverick 128 experts), Scout supports a 10 million context length and fits on one Nvidia H100 for inference, and the family is natively multimodal and 12‑language multilingual. Meta ships Llama Stack, on‑device executor optimizations, Torch/TorchTune fine‑tuning workflows, and a Lamagard classifier pipeline for safety; Hugging Face downloads exceed one billion and 200,000 derivative models exist.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 1:03LLMs and Meta AI
- 2:29Llama model evolution
- 3:34Open source impact
- 4:59Llama capabilities
- 6:47Meta's open source AI
- 8:09Llama Stack tools
- 9:55Llama 4 overview
- 12:29Llama 4 performance
- 13:11Model evaluation
- 13:55Responsible AI
- 15:43Resources
related talks
12mUnpack the Latest Llama 4 Herd
38mMeta's Roadmap for Full Stack AI
11mIntroducing Llama API: A Seamless Way to Revolutionize Your Workflow
17mThe State of the Llama Ecosystem
31mBuilding LLaMA: Meta's Director of GenAI
12mKeynote: The Llama Ecosystem: Past, Present and Future
48mBringing Llama 3 to Life
20mSelf-hosting AI LLMs - a beginners guide