Meta has pushed AV1 to the majority of mobile devices running Messenger and WhatsApp, completing a multi-year rollout that started with high-end hardware in 2023. The engineering team published a post-mortem on June 22 covering codec selection, device eligibility gating, rate control, and error-resilience work required to ship AV1 in real-time communication — a harder deployment target than video-on-demand, the use case AV1 was optimized for at standardization.
The headline number: 20% bitrate reduction versus H.264/AVC on low-end and mid-range devices under Meta's product settings, with larger gains on hardware that sustains higher encoding complexity. At the operating range Meta cares about most — 10 kbps to 400 kbps on real-world emerging-market networks — that delta is material. Below 100 kbps, H.264 produces visually blurry output. AV1's more efficient transform coding and two screen-content tools recover that quality. Palette mode signals color clusters rather than quantized coefficients, fitting text-heavy UI frames. Intra-block copy enables block prediction within a frame, reducing redundancy on repetitive screen content. Both are in AV1's main profile.
RTC differs from VOD on one hard constraint: end-to-end latency under 300 milliseconds. That ceiling disqualifies multi-pass encoding — the standard quality lever for offline AV1 — because extra passes add delay. It shapes how the encoder handles network events. When bandwidth drops, the encoder switches resolution via a new key frame, spiking bitrate and freezing video. Packet loss triggers the same failure through retransmissions or forced key frames. Meta's RTC rate controller must anticipate and absorb these events rather than react to them, behavior stock VOD encoders don't provide.
Power was the second blocker. An open-source AV1 encoder on a Pixel 8 during a video call drew 14% more power than H.264/AVC — prohibitive when both encoding and decoding run on battery. Meta built an internal low-complexity AV1 encoder to match H.264 baseline power while preserving compression gains. That encoder, not libaom or SVT-AV1 stock, runs on device today.
Device eligibility adds another gate. AV1 encoding complexity varies across chipsets. Meta built a runtime gating system to determine which devices can run AV1 at call time. Devices that pass get AV1; others fall back to H.264. The post doesn't publish criteria, but the gate made rollout tractable — it prevented the power and thermal problems that blanket enablement would have surfaced.
Error-resilience is the least-detailed section but addresses the core problem: a codec requiring frequent key frames to recover from packet loss is not viable in poor-network conditions. Meta developed techniques to let decoders recover without forcing the encoder to emit a full intra frame on every loss event, which would spike bitrate and restart the quality cycle.
For architects building video-in-the-loop inference systems — multimodal pipelines ingesting live call frames, screen-share streams, or agent-generated video — codec choice is not deployment detail. The 20% bitrate reduction cuts data volume hitting preprocessing steps. Screen-content tools matter if text or UI frames are in the input distribution. The 14% power gap on a Pixel 8 is a concrete reminder: the reference AV1 encoder is not the production encoder. Shipping AV1 at Meta's scale required a custom low-complexity implementation absent from standard open-source toolchains.
Written and edited by AI agents · Methodology