Engineering at Meta

Engineering at Meta is a technical news resource for engineers interested in how we solve large-scale technical challenges at Meta.

[NVIDIA Newsroom @nvidianewsroom]

NVIDIA introduces the OpenShell Reference Architecture to help businesses run AI agents under control.

The open design combines accelerated compute for agents with hardware-isolated security for deployment at scale.

Announced at GTC Berlin: https://t.co/j4esokYQGF

[NVIDIA @nvidia]

Jensen introduces the OpenShell Reference Architecture to help businesses set boundaries for AI agents and stop them when they step outside those boundaries.

Securing agents around the world. 🔒

[Engineering at Meta @Meta_Engineers]

🎯 Excited to share how we scaled the 2026 generation of GEM — Meta's Generative Ads Model — the central model and compute engine behind ads across Meta.

Last year we introduced GEM, the industry's largest recommendation systems foundation model. This year, we are pushing the frontier again with breakthroughs across model architecture, cross-surface cross-domain learning at Meta’s scale, innovative intelligence transfer, and LLM-scale AI infrastructure:

Bending the scaling curve with Kunlun. GEM's new LLM-style architecture, co-designed with the latest GPU hardware, delivers 2.5x model complexity growth over earlier versions of GEM and processes 2x longer engagement history per person, while modeling the massive feature interactions unique to recommendation systems.

Expanding learning across Meta's ads ecosystem. GEM is the first model to jointly predict clicks and conversions across all surfaces — Facebook, Instagram, Reels — and across major deep-funnel objectives like off-site conversions and mobile app installs.

Doubling the effectiveness of intelligence transfer with LoopFM. This new frontier post-training technique passes GEM's internal representations and historical memory to the models that serve ads as a sequential context window, at no additional compute cost or latency. Elastic pre-computation (Solaris) further harvests gains from upstream models.

Scaling frontier training efficiently. Through deep model and infrastructure co-design, we increased total training FLOPs by 4x while doubling end-to-end training efficiency to 20–25% MFU over the past 12 months, across thousands of latest-generation GPUs.
Recommendation systems don't scale the way language models do — heterogeneous features, shifting distributions, and no option to serve a frontier model in a live request. Making GEM scale like LLM took rethinking the scaling architecture, the post-training and the training stack together. Very proud of this team.

📈 The payoff: for people, ads that better match what they're actually looking for. For advertisers, an 18.9% lift in conversions and 12% in click-through rate across our ad surfaces and objectives over the past year — with 5.8%, 4.1% and 1.6% conversion uplift on Facebook, Instagram and off-site conversion for targeted users.

🔗 Read the full engineering story: https://t.co/T9pbVlNYWj
#AI #MachineLearning #AdsRanking #MLEngineering #Engineering

[AI at Meta @AIatMeta]

We trained our Muse foundational models, including Muse Wonder, at Prometheus, our gigawatt-scale supercluster in New Albany, Ohio.

Across a supercluster with hundreds of thousands of GPUs, even one underperforming host can slow an entire training run. We engineered Prometheus for resilience at scale, achieving 85%+ goodput on pretraining, 90%+ on reinforcement learning and ~47K GPU hours mean time between interruptions during Muse Wonder training.

Take a look at the tech powering Prometheus: https://t.co/WSFODnGRST

[Engineering at Meta @Meta_Engineers]

We're open-sourcing Rebalancer, the assignment-problem solver Meta has used for resource allocation across our infrastructure for over nine years.

Given a set of objects and a set of bins, Rebalancer assigns one to the other to optimize your objectives under your constraints, including hardware placement, service and task placement, traffic routing, and more. It separates how a problem is specified from how it's stored, solved, and debugged, with both an optimal MIP solver and a parallelized local search solver.

At Meta it solves ~40 million assignment problems a day across 30+ problem formulations. P99 solve time is 12 seconds on a problem with 265k objects and 3.2k bins.

🔗 Read more on our engineering blog: https://t.co/9D3jAGfOdQ

[Engineering at Meta @Meta_Engineers]

Impressive results from a great collaboration between our teams! Read the full deep dive 👇

[Android Developers @AndroidDev]

The Instagram Direct engineering team modernized @Instagram messaging at scale using Jetpack Compose. 🏗️

By replacing their legacy architecture with an AI architecture and Compose-native alternatives, they maintained a high performance bar and reduced their UI codebase by 50%. 🤯

Read more → https://t.co/0YwtzWT0LA

[Meta for Developers @MetaforDevs]

Muse Spark 1.3 is now available through @Oracle, with availability on @GoogleCloud in private preview.

We’re excited to give developers more choice in where they build and deploy by expanding access to our frontier coding and agentic model through our preferred cloud platform partners.

[Engineering at Meta @Meta_Engineers]

Last year we introduced Private Processing, Meta’s confidential computing infrastructure that keeps everything you share with your AI assistant private, so that no one, not even Meta, can access your data.

At #MetaConnect, we announced that we’re bringing these same privacy benefits to our AI glasses.

Read the technical deep dive on our approach: https://t.co/hjYAndwu3e

[Engineering at Meta @Meta_Engineers]

Today we’re announcing Petal: the world’s first petabit-class transoceanic subsea cable and the first to deploy multi-core fiber at scale.

Spanning 7,000 km (4,000+ mi) between France and the United States, Petal will deliver 1 Petabit per second (1 Pbps / 1,000 Tbps), doubling the capacity of today’s most advanced transatlantic systems.

Petal is set to enter service in 2029, setting a new standard for high capacity, high density SDM subsea cable designs.

Read the full technical breakdown on our engineering blog: https://t.co/Ksn9oM9gBA

[Engineering at Meta @Meta_Engineers]

To help Meta subject matter experts focus more on work where their knowledge matters most, we’ve built an AI agent that acts as a secondary expert for a given domain. This system makes deep specialist knowledge readily available and preserved for anyone in an organization to access, share, and build upon.

Two things make it novel and portable:

1️⃣ structured, auditable knowledge architecture that separates what the agent knows from how it reasons

2️⃣ self-improvement loop that turns expert feedback into verified updates without retraining

We’re sharing a deep dive into our approach and architecture here: https://t.co/oqijON6ztE

[Engineering at Meta @Meta_Engineers]

How do you manage database connections when you have 1M+ clients hitting 500k+ database hosts? We’re sharing a deep dive into ZGateway, the stateless proxy tier handling traffic for ZippyDB (>1B ops/sec).

The direct-access model created an unbounded TLS connection mesh, wasting CPU and causing OOM crashes during reconnection storms. ZGateway decoupled the fleets, dropping per-host connections by 97–98% with just a 6% compute overhead.

Read the full engineering deep dive here: https://t.co/qKtPDJBND4

[Muse @Muse]

Introducing Muse, your personal AI agent from Meta that gets things done across every part of life.

Download the Muse app and get started: https://t.co/KBjYWfshGo

[Mark Zuckerberg @finkd]

Introducing Muse, the personal agent that understands your goals and works 24/7 to get things done for you.

[Engineering at Meta @Meta_Engineers]

Training and serving frontier AI models depends on fast, reliable networks that move data between GPUs without wasting compute cycles.

To meet this challenge at scale, we designed MetaRoCE: a clean-sheet RDMA transport protocol purpose-built for AI workloads on commodity Ethernet.

We’re releasing the MetaRoCE specification, a reference software implementation and a compliance test suite through the @OpenComputePrj to enable the broader industry to adopt, implement, and build on it.

Read the full technical deep dive: https://t.co/VXGLrRc2v8

[Engineering at Meta @Meta_Engineers]

Read more about the entire MTIA family: https://t.co/nYqeRGxUbC

[Engineering at Meta @Meta_Engineers]

By now you’re already familiar with MTIA, our family of in-house training and inference accelerators.

Today, we’re sharing a deep dive into MTIA 300, the first MTIA chip optimized for training ranking and recommendation models. MTIA 300’s built-in NIC chiplets allow it to meet the communication needs associated with training recommendation models with superior performance over general-purpose GPUs.

See the full technical deep dive from the @ Scale: Networking Conference: https://t.co/ajpH6s5Ryr

[Engineering at Meta @Meta_Engineers]

The convergence of sequence modeling and recommendation systems has opened new frontiers in artificial intelligence.

🚀 Our new multi-layer sequence learning architecture demonstrates LLM-like scaling law, while delivering performance in latency-sensitive ads ranking models. This new architecture is deployed for user representation within Meta’s ads recommendation as a resource efficient, multi-stage system that boasts high upstream-to-online transfer ratios.

👉 Read our research paper to learn more: https://t.co/SJVKz312W8

[Engineering at Meta @Meta_Engineers]

Meta’s Generative Ads Recommendation Model (GEM), the engine behind ads on Instagram and Facebook, is now training at LLM scale across thousands of the latest-gen GPUs.

Because standard LLM infrastructure doesn't natively fit recommendation workloads, we co-designed custom kernels like Jagged Flash Attention (JFA), Generalized Dot-Product Attention (GDPA), BlockAttention, ultra-low mixed precision (including MXFP8 attention), and a topology-aware 5D parallelism architecture tailored to our network hierarchy.

The result: we scaled total training FLOPs 4x in 12 months while doubling end-to-end training efficiency to 20–25% MFU.

Read the full deep dive here: https://t.co/btMBljsNVS