Meta Ditches Its Private Protocol and Ships Instagram Live on Open MoQ

Meta has deployed Media over QUIC (MoQ) as the primary ingest transport for Instagram Live broadcasts from iOS and Android, according to its engineering report. MoQ replaces Meta’s in-house RUSH protocol on the publisher-to-edge path, which carries video from a creator’s device into Meta’s network. An edge translation layer preserves the existing backend and playback systems, making the change largely invisible to creators and viewers.

An open protocol replaces a private one

For nearly a decade, Meta used Reliable Ultra-low-latency Streaming over HTTP (RUSH), its proprietary QUIC-based successor to RTMP, for live-video ingest. Maintaining a company-specific protocol limited outside implementation, interoperability testing, and shared development. MoQ gives Meta a protocol shaped through the Internet Engineering Task Force, with contributions from companies including Google, Akamai, Cisco, and Cloudflare, as well as academic researchers.

Live-video services commonly use separate technologies for ingest, large-scale playback, and interactive communication. RTMP often handles ingest, HLS or DASH distributes buffered video, and WebRTC supports calls and other low-latency sessions. MoQ aims to provide a shared object model and QUIC-based transport across those workloads, reducing the number of protocol stacks and operational systems a platform must maintain.

MoQ remains an IETF draft, so implementers should expect wire details and APIs to evolve. Meta’s deployment provides production experience that can inform the specification while exposing interoperability and operational issues at large scale.

Tracks, objects, and relays

MoQ packages media into addressable objects organized as tracks. A publisher sends those objects through relays, and subscribers request the tracks they need. Relays use MoQ metadata to cache, prioritize, forward, or discard objects while treating encoded audio and video payloads as opaque data. The resulting architecture resembles a content delivery network with media semantics built into the transport.

• Independent delivery: QUIC can carry media objects on separate streams, preventing packet loss in one stream from stalling unrelated streams at the transport level.
• Subscriber-specific latency: Subscribers to the same track can use different buffering targets. An interactive participant might target roughly 100 milliseconds, while an archival service can accept several seconds.
• Object-aware congestion handling: Relays can prioritize recent or essential objects and abandon stale data when bandwidth drops, instead of treating the session as one undifferentiated byte stream.
• TCP-compatible fallback: Deployments can preserve MoQ semantics through mappings such as WebTransport over HTTP/2 or the emerging QMUX work when networks block UDP on port 443.

Migration without a backend rewrite

Meta limited the initial deployment to the ingest leg and divided the migration into client publishing, edge translation, and server-controlled routing. That structure allowed engineers to compare MoQ with RUSH, roll traffic back quickly, and leave the mature downstream video pipeline intact.

1. Client publishing: Meta built the publisher with moxygen, its open-source MoQ library, and integrated it with the company’s mvfst QUIC stack. A configuration flag selected MoQ or RUSH for each broadcast.
2. Edge translation: Dedicated endpoints terminated MoQ connections and converted incoming frames into the RUSH format expected by the existing backend.
3. Server-controlled routing: Stream URL generation directed eligible broadcasts to MoQ-capable endpoints. Once dual-stack clients reached users, Meta could adjust the rollout without requiring another app release.

Network compatibility required additional safeguards because enterprise firewalls and carrier networks can block or rate-limit UDP on port 443, which QUIC normally uses. Meta races QUIC and TCP connection attempts and selects the path that succeeds first. QUIC encryption also prevents conventional packet captures from exposing MoQ framing, so engineers added edge tracing that records object-level events at the protocol endpoint.

Parity clears the migration gate

Meta reports that MoQ matched RUSH across the streaming-quality and reliability indicators used during the rollout. Matching the incumbent allowed the team to change the transport while holding creator and viewer performance steady. The edge translator also isolated failures from the rest of the production pipeline and provided a practical rollback path.

The result establishes that MoQ can handle Instagram Live’s mobile ingest workload, although it does not yet demonstrate an end-to-end MoQ media system. Delivery to viewers, backend processing, and playback remain on Meta’s established infrastructure.

Playback and interactive AI come next

Extending MoQ from ingest through playback would let Meta use object priorities and partial reliability across the full media path. A player could abandon stale video after packet loss, limiting accumulated delay, while adaptive bitrate logic could lower resolution to preserve continuity under constrained bandwidth.

Meta is also exploring MoQ for real-time interactive AI. A bidirectional transport capable of carrying media and signaling at low latency could support conversational agents that continuously exchange audio, video, and model output without requiring separate protocol stacks for each channel.

What developers can use now

• An open client implementation: Moxygen provides reusable MoQ components and integration points for custom media handlers.
• A public specification: The IETF transport draft documents the evolving protocol. Implementations should pin and report the draft version used for interoperability testing.
• A low-risk migration pattern: Terminating MoQ at the edge and translating into an existing internal format allows teams to evaluate a new transport without replacing the entire media pipeline.
• Production evidence with defined limits: Instagram Live validates mobile ingest at Meta’s scale. End-to-end delivery, cross-vendor interoperability, and application-specific performance still require separate testing.