Meta's Reality Labs ported Android's Cuttlefish virtual device framework to macOS on Apple silicon, consolidating a split Goldfish and Cuttlefish setup into one high-fidelity platform. The post details the technical changes: making QEMU's vhost-user path portable beyond Linux, increasing memory-region tracking slots in the HVF backend, adding a native Cocoa display, and handling JIT entitlements and code signing.
Meta engineering blog
Loading
Meta's engineering team describes Intelligent Delivery, an ML-driven system that allocates compute across its ads serving stack. The write-up covers a Resource Core Model that predicts the marginal utility of compute per request, a hierarchy of decision engines spanning regions, hardware, and host-level scheduling, and agent-driven optimization loops. The authors report about 10% incremental ranking gains from 2017 through the first half of 2026.
Meta's engineering team describes Your algorithm, a suite of LLM-powered tools that lets Facebook users edit their interests and exclusions to steer Feed and Reels recommendations. The post details a four-source retrieval design for new interests, layered filtering for Not Interested In topics, and a small fine-tuned reranker that removes matching candidates. It also explains using a separate LLM as a judge to evaluate prompts and models offline.
Meta describes TLX, an extension to Triton that adds explicit warp-group orchestration, synchronization via mbarriers, TMA and TMEM access, and cluster-level execution for GPUs. The post explains the 'missing middle' programming model (MIMW) between CUDA's SIMT and Triton's blocked model, and reports production results in GEM training (20-25% E2E MFU), RecSys inference (IKBO, up to two-thirds net-latency reduction), and distributed grouped GEMM for LLM pretraining.
Meta describes a vendor-neutral accelerator plugin architecture in lldb-server that lets LLDB debug applications spanning CPUs and GPUs. The design models the CPU and GPU as separate debugger targets coordinated by a plugin that uses internal breakpoints and structured protocol actions, avoiding a forced unified process model. The post notes that most of the common framework has been upstreamed to LLVM, with a mock plugin enabling hardware-independent testing.
Meta describes F3 Lightning, an infrastructure layer that decouples feature releases in its ads ML feature framework so each feature can be compiled, validated, deployed, and rolled back independently. The team preserved shared-execution efficiency through stable computation identifiers, a single flattened execution plan, a shared per-request memory buffer, and dense index mapping. The result is comparable features reaching production about 3.6x faster with no increase in compute cost.
Meta describes the production deployment of Catching Just-in-Time Tests (JiTTests), which use LLMs to generate executable tests from code diffs before they land. The post covers two complementary generation workflows (intent-aware and dodgy diff), a false-positive filtering pipeline combining LLM-as-Judge ensembles with a deterministic rule-based assessor, and a human-on-the-loop review model. Reported results include a 60% true-positive rate and a 20% reduction in privacy SEVs in the first...
Meta describes the next generation of its Andromeda ads retrieval engine, which uses a retrieval mixture-of-experts architecture, a single fused GPU kernel, and co-located index matching to score every eligible ad at over one billion ads per second. The post details network and memory bottlenecks encountered at billion-scale throughput and the co-design fixes applied, along with NVIDIA kernel optimizations yielding 2x to 5x speedups.
Meta's engineering team describes PrivacyLib, a common library that runs inside the clients of hundreds of data and processing systems to enforce privacy requirements at the point of data access. The post covers how it separates system integration from requirement modules, how it keeps billions of checks per second inexpensive, and how enforcement changes are simulated and rolled out safely. It also explains how the same integration extends to observing and gating internal AI agent tool calls.
Meta's 2026 Python Typing Survey, with 1,794 responses, finds stable type hint adoption but a sharp shift in type checker choice toward newer tools. Missing third-party library type information is the top difficulty, and half of respondents now use more than one checker. The survey also examines how AI coding tools interact with type checkers, a question the authors say still needs empirical study.
Meta outlined the networking work behind its latest AI models and gigawatt-scale Prometheus cluster at the 2026 OCP Global Summit. Highlights include the liquid-cooled IcePack/IceCube 102.4T switch rack running FBOSS, a flattened multi-plane scale-out fabric, the BAG super-spine for scale-across, Ethernet scale-up with Tahan, and the open-sourcing of MetaRoCE. The post also indexes the accompanying talks and slide decks.
Meta describes Prometheus, its gigawatt-scale GPU supercluster in New Albany, Ohio, covering the GB200 and GB300 NVL72 platforms, the Ethernet back-end fabrics, and the reliability work that cut interruptions by 99%. The post details fault recovery through parallelism-aware fault tolerance and checkpointing, along with power transient management and liquid cooling tradeoffs.
Meta describes GenRecs Tokenization, which represents ads, products, and pages as short sequences of discrete semantic tokens instead of dedicated IDs, so new items can inherit signal from related items. The post explains a three-stage pipeline of embedding, subspace decomposition, and codebook quantization, and links intrinsic metrics such as cluster purity, distribution balance, and codebook utilization to downstream CTR accuracy. It reports a +0.35% conversion lift in production.
Meta's engineering team describes a 'thin orchestrator' pattern for their AI data-pipeline agents, where a deterministic router handles sequencing and reliability gates while 14 specialized subagents perform domain reasoning. The post reports that a single-agent approach hallucinated under heavy context, and that workflow templates and runtime-loaded knowledge files cut column configuration time from days to under an hour. Practical guidance is offered for separating routing from reasoning and...
Meta describes how a small team built Forum, a standalone Facebook Groups app, by treating an AI agent as a participant in the engineering cycle. The post covers a shift from human-in-the-loop to human-on-the-loop workflows, verification-first quality gates, build-level isolation for prototypes, and agent-driven stress testing that found and fixed a memory issue. Reported outcomes include reaching feature parity with Facebook Groups in eight weeks and roughly halving memory use and page faults.
Meta's engineering team describes how its infrastructure absorbed record 2026 World Cup traffic, including WhatsApp peaks of over 30 million messages per second. The post explains the capacity buffers, country-level graceful degradation, cascade prevention, and region-specific forecasting used to handle synchronized global spikes without a deployment freeze.
Meta engineers Ke Mao, Mark Harman, and Yue Jia received the ISSTA 2026 Most Influential Paper Award for their 2016 Sapienz research on multi-objective automated testing of Android apps. The post describes how Sapienz balances app coverage, crash discovery, and reduced reproduction steps, and reports its scale at Meta, where it runs millions of tests daily. It also notes how the approach is being extended toward LLM agents for functional correctness testing.
Meta describes the 2026 generation of its Generative Ads Model (GEM), a central RecSys foundation model that unifies ads prediction across surfaces and objectives. The post covers the Kunlun LLM-style architecture with block attention, GDPA kernels, and CompSkip for ultra-long sequences, plus LoopFM and Solaris for transferring knowledge to serving models. It also details training at LLM scale, reporting 4x total FLOPs and doubled MFU to 20-25% across thousands of GPUs.
Meta explains how it deployed Media over QUIC (MoQ) for Instagram Live ingest on iOS and Android at production scale. The post covers MoQ’s track-based architecture, QUIC transport benefits, edge translation, phased rollout controls, TCP fallback, observability, and parity with the existing RUSH protocol.
Meta describes an inference-serving system that converts fragmented, preemptible idle capacity into better ads recommendations. Its Rank More Ads and Serve Better Models levers, isolated routing, feedback control, and millisecond-scale value-aware load shedding improve utilization while protecting baseline latency and reliability.