AI at Meta

📣 Today we're opening a call for applications for Llama 3.1 Impact Grants! Until November 22, teams can submit proposals for using Llama to address social challenges across their communities for a chance to be awarded a $500K grant. Details + application ➡️ https://go.fb.me/rd22jf This year we're expanding the Llama Impacts Grant program by hosting a series of virtual events and in-person hackathons, workshops and trainings around the world — and providing technical guidance and mentorship to prospective applicants. These programs will support organizations in Egypt, Hong Kong, India, Indonesia, Japan, the Kingdom of Saudi Arabia, Korea, Latin America, North America, Pakistan, Singapore, Sub-Saharan Africa, Taiwan, Thailand, Turkey, the United Arab Emirates and Vietnam. We’re inspired by the diverse projects we’ve seen developers undertake around the world to positively impact their communities by building with Llama and we're excited to support a new wave of global community impact with the Llama 3.1 Impact Grants.

📣 New and updated! Try experimental demos featuring the latest AI research from Meta FAIR! • Segment Anything 2: Create video cutouts and other fun visual effects with a few clicks. • Seamless Translation: Hear what you sound like in another language • Animated Drawings: Bring hand-drawn sketches to life with animations. • Audiobox: Create an audio story with AI-generated voices and sounds. Try the research demos ➡️ https://go.fb.me/brn8mg

The MLCommons #AlgoPerf competition was designed to find better training algorithms to speed up neural network training across a diverse set of workloads. Results of the inaugural competition were released today and we’re proud to share that teams from Meta took first place across both external tuning and self-tuning tracks! 🔗 Details • Results from MLCommons ➡️ https://go.fb.me/poejsh • Schedule Free ➡️ https://go.fb.me/5wf35d • Distributed Shampoo research paper ➡️ https://go.fb.me/tns64m

📣 Just announced by Mark Zuckerberg at SIGGRAPH! Introducing Meta Segment Anything Model 2 (SAM 2) — the first unified model for real-time, promptable object segmentation in images & videos. In addition to the new model, we’re also releasing SA-V, a dataset that’s 4.5x larger + has ~53x more annotations than the largest existing video segmentation dataset in order to enable new research in computer vision. Details ➡️ https://go.fb.me/edcjv9 Demo ➡️ https://go.fb.me/fq8oq2 SA-V Dataset ➡️ https://go.fb.me/rgi4j0 SAM 2 is available today under Apache 2.0 so that anyone can use it to build their own experiences. Like the original SAM, SAM 2 can be applied out of the box to a diverse range of real-world use cases and we’re excited to see what developers build.

The livestream from #SIGGRAPH2024 kicks off at 3pm PT today ⬇️

As part of our release of Llama 3.1 and our continued support of open science, this week we published the full Llama 3 research paper that covers a range of topics including insights on model training, architecture and the results of our current work to integrate image/video/speech capabilities via a compositional approach. The Llama 3 Herd of Models Paper ➡️ https://go.fb.me/1nmc78 We hope that sharing this research will help the larger research community understand the key factors of foundation-model development and contribute to a more informed discussion about the future of foundation models in the general public.

Ready to start working with Llama 3.1? Check out all of the newest resources in the updated repos on GitHub. Official Llama repo ➡️ https://go.fb.me/wy18hm  Llama recipes repo ➡️ https://go.fb.me/95swt9 Llama Reference System ➡️ go.fb.me/q0cjcb The repos include code, new training recipes, an updated model card, details on our latest trust & safety tools, a new reference system and more.

Llama 3.1 is available right now across many of our partners! Ready to go? Here’s a quick look at features across all currently available public API pricing as of this afternoon. Details on pricing, capabilities and links to where you can get started will be updated here ➡️ https://go.fb.me/3qft64

Starting today, open source is leading the way. Introducing Llama 3.1: Our most capable models yet. Today we’re releasing a collection of new models including our long awaited 405B. Llama 3.1 delivers stronger reasoning, a larger 128K context window & improved support for 8 languages including English — among other improvements. Details in the full announcement ➡️ https://go.fb.me/hvuqhb Download the models ➡️ https://go.fb.me/11ffl7 We evaluated performance across 150+ benchmark datasets across a range of languages — in addition to extensive human evaluations in real-world scenarios. Trained on >16K NVIDIA H100 GPUs, Llama 3.1 405B is the industry leading open source foundation model and delivers state-of-the-art capabilities that rival the best closed source models in general knowledge, steerability, math, tool use and multilingual translation. We’ve also updated our license to allow developers to use the outputs from Llama models — including the 405B — to improve other models for the first time. We’re excited about how synthetic data generation and model distillation workflows with Llama will help to advance the state of AI. As Mark Zuckerberg shared this morning, we have a strong belief that open source will ensure that more people around the world have access to the benefits and opportunities of AI and that’s why we continue to take steps on the path for open source AI to become the industry standard. With these releases we’re setting the stage for unprecedented new opportunities and we can’t wait to see the innovation our newest Llama models will unlock across all levels of the AI community.

We recently announced Joint Audio and Symbolic Conditioning for Temporally Controlled Text-to-Music Generation — or JASCO. This model for music-generation differs from previous approaches with improved controllability using conditioning inputs like chords or beat. We've shared more details on this work for the community in the research paper ➡️ https://go.fb.me/sou2ox