AI at Meta

In the past 6 months... - Muse Spark - Muse Image - Muse Video - Muse Spark 1.1 - Meta Model API - Muse Glimmer - Muse Spark 1.2 - Muse Code - Muse Spark 1.3 - Muse ...and we’re just getting started. 🚀

Mark just unveiled Muse Realtime Avatar at #MetaConnect. Muse Realtime Avatar is our state-of-the-art embodiment technology that turns Muse Realtime Voice into expressive, interactive avatars. This technology enables an entirely new range of interactions in Muse, starting with real-time conversations. Learn more on our research blog: https://lnkd.in/e6bvpqXw

Today we’re launching Muse for Mac, the first version of Muse that can get things done directly on your computer. It works with your files, Messages, Calendar, Notes, Mail, and more, where they already live. Key features: —> It's designed for everyday tasks across your apps: organizing folders, building an end-of-day summary from email, chat, and notes, or completing a half-finished form using files already on your Mac. —> Access is opt-in and yours to control. Full Disk Access is your choice to enable. Sensitive actions, like deleting files or sending messages, always require your approval first. Permissions can be changed anytime in Settings. More in this space coming soon! Try it today at ai.meta.com/muse/download/

Today the FAIR Chemistry team, in collaboration with Andrew Ferguson's group at the Pritzker School of Molecular Engineering at the University of Chicago demonstrated that we can simulate entire enzymatic reactions with explicit solvent (O(100k) atoms), using the reactive potential eSEN-omol (part of the UMA model family), at near-quantum accuracy, matching real experimental measurements for enzyme activity, about 1,000x times faster compared to SOTA QM/MM methods, with no fine-tuning or system-specific setup associated with QM/MM. We demonstrate strong generalization across a variety of enzyme catalysis examples and show that the model works out of the box for any enzyme or biological system and can easily be applied to full chain enzymes in solvent with as little as 8 GPUs, with simulation speeds of millions of steps per day with additional parallelism. Enzyme engineering has the potential to revolutionize many domains such as sustainable plastic recycling, drug discovery, and cleaner manufacturing. Novel generative protein language models can help ideate enzyme patterns but cannot measure enzyme activity or the fundamental reactions that govern enzyme catalysis. Simulating this complex process requires long-timescale dynamics with quantum mechanical accuracy, which is enormously expensive. Up until now, the community has used shortcuts (such as QM/MM or ML/MM) to study smaller groups of atoms using these quantum methods and classical force fields to approximate the rest. By demonstrating the ability to use an MLIP to simulate the entire system at scale via a DFT surrogate, we open an entirely new realm of science for computational enzymology and biology. To enable this work, we worked together with the PyTorch team (special thanks Aditya Venkataraman) at Meta to make the UMA line of models significantly faster, reaching speeds exceeding 3ns/day (3 million steps/day) for 1000 atoms on a single H200 GPU. To enable UMA to simulate at scales of biomolecules, we seamlessly integrated distributed inference that allow anyone to scale UMA to any number GPUs with a simple config change, allowing bio simulations with 100k+ atoms to run at ns per day speeds. The latest versions of the accelerated UMA and distributed inference framework is fully available in the FAIR Chemistry codebase. We like to thank our incredible collaborators and the Fairchem team: Armin Shayesteh Zadeh, Aniruddha Seal, Siva K. Dasetty, PhD, Siddarth Achar, Ph.D., Misko Dzamba Benjamin Miller, Leif Jacobson, Larry Zitnick, Brandon Wood, Zachary Ulissi, Daniel Levine, Andrew Ferguson Paper: https://lnkd.in/gZK4CaAb Fairchem with code and free access to all models: https://lnkd.in/gHheZmKF

Today we launched Muse, a personal AI agent, and published a deep dive on how we built safety into its system. An agent that gets to know you over time necessarily holds a lot of context about you. That's the source of its usefulness and the reason we built Muse to be secure, safe, and private. Read the full deep dive here: http://security.muse.ai/

Introducing Muse, a personal agent that gets things done for you, powered by Muse Spark 1.3. Get an inside look at how we built Muse and what it can do for you: https://lnkd.in/erNjfjwC

Today we’re excited to release Muse Spark 1.3 with improved performance on agentic and coding tasks, and a focus on real-world usability. Key capabilities: 1️⃣ Long-horizon agentic tasks: generates its own context across conflicting sources, self-corrects plan gaps, tracks what it learned 2️⃣ Real collaboration: asks clarifying questions, invokes help when stuck, confirms before consequential actions 3️⃣ Multitasking: maps incoming prompts to the right task inside a single, messy thread — whether you're steering an earlier request or interrupting it 4️⃣ Coding efficiency: ~20% fewer tool calls and ~25% fewer tokens vs. 1.2, per internal comparisons 5️⃣ Safety: stronger adversarial robustness and better calibration on irreversible actions Muse Spark 1.3 is rolling out today in Muse Code and Meta Model API at dev.meta.ai, with max reasoning coming soon after we finish safety testing. We also have bigger models, Muse Spark open weights, and more on the way soon. Learn more: https://go.meta.me/46IsHsV

Introducing Muse Voice Transcribe, the first real-time audio perception model developed by Meta Superintelligence Labs. Muse Voice Transcribe delivers real-time streaming ASR, diarization with 20+ speakers, and endpointing. It’s multilingual with seamless code-switching and improves accuracy with language, keyword, and context biasing. The model ranks first on Artificial Analysis streaming speech-to-text and on public diarization benchmarks. Muse Voice Transcribe is available today via Meta Model API, Meta AI for Mac, and Muse Code. Learn more: https://go.meta.me/4xISnl2