Engineering at Meta

This year marks Meta’s 10th consecutive year as a sponsor of the Python Software Foundation!

Python is the most heavily used programming language across Meta’s engineering stack. It serves as the architectural backend for core products like Instagram and Threads, and acts as the foundation for our cutting-edge AI research and infrastructure.

Our decade-long technical investment focuses on ensuring the open-source ecosystem remains performant, secure, and sustainable. Key areas of our ongoing collaboration and contribution include:

1️⃣ Core Maintenance: Meta engineers serve as core maintainers of the language, directly authoring new features and Python Enhancement Proposals.

2️⃣ Open-Source Tooling: We build open-source Python developer tools to help developers write better quality, more performant Python, including projects like Pyrefly (a fast type checker and language server).

3️⃣ Community Growth: Our support helps fund educational programs and community events like PyCon US and PyLadies to grow the community and foster new talent.

Read our full reflection on our long-term commitment to Python: https://engineering.fb.com/2026/06/30/open-source/10-years-of-metas-commitment-to-python/

Deploying advanced video codecs for real-time communication (RTC) at a global scale introduces constraints across latency, compute complexity, and device heterogeneity.

We have successfully adopted the AV1 video codec across the majority of mobile devices running Meta RTC applications such as Messenger and WhatsApp.

Under product settings on low-end and mid-range devices, AV1 achieves a minimum 20% bitrate reduction compared to H.264/AVC, crucial for maintaining high visual quality below 100 kbps on limited networks. Additionally, AV1’s main profile palette mode and intra-block copy tools significantly improve complex screen content and text compression.

Overcoming the operational and technical challenges of production deployment required cross-layer optimization:

1️⃣ Encoder & Decoder Selection: To mitigate a 14% power consumption overhead observed with open-source AV1 encoders on mobile devices, we implemented an internal low-complexity encoder preset that aligns with H.264 baseline power metrics. For decoding, we integrated the dav1d decoder to optimize power efficiency and mitigate audio/video desynchronization.

2️⃣ Binary Size Mitigation: App-size expansion affects metrics like memory usage and crash rates. By stripping unused tools—such as the quantization matrix (QM)—from the end-to-end pipeline, we reduced library overhead by 60 kB.

3️⃣ ML-Driven Device Eligibility: Traditional hardware categorization rules are unreliable across millions of distinct Android configurations. We developed a machine learning framework that evaluates low-level production performance metrics to yield a dynamic rtc_score. Iterating from Model V1.1 to a two-tier Model V2 allowed us to safely maximize AV1 enablement across entry-level to flagship architectures.

Read the full deep dive to learn more: https://engineering.fb.com/2026/06/22/video-engineering/adopting-av1-for-real-time-communication-rtc-meta/

For years, recommendation retrieval systems have relied on a complex mesh of microservices. But as scale and sophistication grew, so did the structural limits like network latency, version inconsistencies, and siloed development environments.

To break through these barriers, we’re introducing SilverTorch, a reimagined retrieval ecosystem built under a new paradigm: "Index as Model."

Instead of stitching neural networks into a microservice architecture, we’ve built the entire retrieval system as a single PyTorch neural network. Every component – the item index, eligibility filtering, scoring, and user tower – is now a tensor or operator inside one unified model.

A few highlights of the impact:
🔹 Up to 23.7x higher throughput than state-of-the-art approaches
🔹 20.9x more compute cost efficiency vs. CPU baselines
🔹 Unlocked capacity for complex neural reranking without breaking our strict sub-100ms latency budget
🔹 Greater engineering velocity by dissolving the boundaries between ML and infrastructure

Read the full breakdown of how we built SilverTorch and details from our SIGIR 2026 paper: https://engineering.fb.com/2026/05/26/ml-applications/silvertorch-index-as-model-new-retrieval-paradigm-recommendation-systems/