Skip to content

CONCEPT Cited by 1 source

Device eligibility (ML-based)

ML-based device eligibility is the approach of using production telemetry and machine learning — rather than heuristic rules (memory, release year, OS version) — to determine whether a specific device model can run a given codec or feature in real-time.

Why heuristics fail

For Meta's AV1 RTC deployment, the challenge was classifying which of the vast number of Android device models (thousands of variants) can encode/decode AV1 in real-time. Meta tried device memory, release year, and Android OS version — none proved sufficiently reliable (Source: sources/2026-06-22-meta-adopting-av1-for-real-time-communication-rtc-at-scale).

Architecture

Meta's ML-based device eligibility framework:

  1. Data collection — Low-level performance metrics (encoding latency, decode timing, power draw, etc.) collected via Meta's logging pipeline from real AV1 calls in production.
  2. Model training — ML model takes device metrics as input features, outputs an rtc_score quantifying AV1 performance capability.
  3. Decision — The score determines whether AV1 is enabled, which encoder preset to use, and other call settings.

Iterative refinement

  • Model V1.1 (August 2025) — Broadened AV1 traffic across more devices, generating dedicated AV1-only training data.
  • Model V2 — Introduced a two-tier approach differentiating higher-end from lower-end devices, reflecting that entry-level phones and flagships have very different encoding capabilities.

The iterative loop is self-reinforcing: more AV1 traffic → more representative data → better model → enables more devices.

Contrast with iOS

iOS device eligibility is straightforward due to Apple's limited number of device variants — simple model-list approach works. The ML framework is specifically needed for Android's device fragmentation.

Seen in

Last updated · 559 distilled / 1,651 read