Measured end to end, reported with the conditions.
We don't publish a number without the run behind it. Every figure below comes from a run whose raw output we keep, and each is labelled with the rig it ran on.
Measured in lab: 6,000 viewers on one 4-vCPU relay, every frame deliveredLab results, not production SLAs. Phone and relay runs on August 2026 builds, and a run through Cloudflare's MoQ relay in October 2026. They don't guarantee performance on your network. More load testing is under way, and this page is updated as results come in.
56–57 ms median, 70 ms p95
Glass to glass is the time from light hitting the sender's camera to the frame appearing on the receiver's screen. It covers capture, encode, packetising, the network both ways through the relay, the jitter buffer, decode and display. That is everything a person on a call experiences, not just the network.
About 22 ms of the median is the network round trip to the relay. Most of the rest happens on the phones.
Measured in lab: 6,000 viewers on one 4-vCPU relay, every frame deliveredRig
| Devices | Two iPhone 15 Pros, chosen as a hard case: a model with known heat issues, older than the two latest generations MOQOM targets. Newer phones are preferred for clients. |
|---|---|
| Relay | Google Cloud Compute Engine, Los Angeles (us-west2) |
| Network | Ordinary home Wi-Fi in Venice Beach, California (about 108 Mbps down, 8 Mbps up), shared with TVs and other devices streaming movies during the runs. Mean round trip to the relay: 21.8 ms |
| Media | One 720p HEVC rung from the camera, hardware encode |
| Direction | Both phones publishing and subscribing at once, 90 s, every frame delivered |
| Clocks | Both synchronised to NTP. The clock-agreement error is reported next to every reading |
Latency by resolution
What is measured so far, and what is supported. How MOQOM compares →
| Resolution | Typical use | Median glass-to-glass | Status |
|---|---|---|---|
| 4K (2160p30) | Solo host, broadcast | Not yet measured | 4K rung in the Swift SDK camera ladder (opt-in) |
| 1080p30 | Host / spotlight | Not yet measured | Supported (top rung of default camera ladder) |
| 720p30 | Two-person call | 56–57 ms (p95 70 ms)* | Measured: two iPhone 15 Pros, home Wi-Fi, relay in Los Angeles, both directions at once, 90 s, every frame delivered |
| 540p60, 600 kbps | Co-host tiles | 69 ms* | Measured, same relay |
| 360p30 | Co-host tiles in shared calls | Not yet measured | Supported (600 kbps rung) |
| Shared video calls (360p–540p per tile) | Group calls, co-host grids | Not yet measured | Closest references: 540p60 at 69 ms; 720p30 with the layout changing every 8 s at 61 ms |
* Built to scale: a Rust media plane and a Go control plane, each scaling out horizontally. Measured in lab: 6,000 viewers on one 4-vCPU relay, every frame delivered. Full data at moqom.cloud/benchmarks/load-tests.md.
6,000 viewers from one 4-vCPU relay, every frame delivered
One publisher, one track and up to 6,000 viewers on a single 4-vCPU machine. Every viewer received every frame, and the relay's own counters were read alongside the viewers' reports, so a relay that quietly gave up on some viewers can't look clean.
The median frame reached viewers in under 40 ms up to 5,500, and in 61 ms at 6,000, with the relay at 97% CPU; the slowest frame took 0.38 s. 6,000 is the largest size we have run on one machine, not the ceiling: relays scale out, so production traffic is spread across as many as it needs.
Measured in lab: 6,000 viewers on one 4-vCPU relay, every frame deliveredRelay: Google Cloud n2-standard-4 (4 vCPU), us-east1-b. Viewers: two n2-standard-8 load generators. Publisher: n2-standard-2. Synthetic traffic, 30 fps of 4,096-byte frames, 30 s per size. One zone, gVNIC, MTU 1460. The relay was started fresh for each size. Latency here is delivery, from the publisher sending a frame to a viewer receiving it. Egress reached 5.46 Gbit/s, about ~1.3–1.4 Gbit/s per vCPU.Full test data →
The same 6,000 viewers on Google Cloud Platform (GCP) and Cloudflare
We ran the same 6,000-viewer test, with the same load generator and traffic, on two networks: Google Cloud Platform (GCP), with MOQOM's relay on a single 4-vCPU machine, and Cloudflare's MoQ relay network, which is in beta. On GCP, all 6,000 viewers received every frame, confirmed by the relay's own counters. On Cloudflare, every viewer who received video got every frame, but 241 got none: 23 were refused when connecting and 218 were accepted but sent nothing.
Cloudflare's global network delivered frames faster than one GCP machine running at 97% CPU. Below full load, the GCP relay's median was 13–17 ms, and in production MOQOM spreads viewers across as many relays as it needs, on GCP or anywhere else.
Measured in lab: 6,000 viewers on one 4-vCPU relay, every frame delivered| 6,000 viewers | Google Cloud Platform (GCP) | Cloudflare MoQ (beta) |
|---|---|---|
| Viewers receiving video | 6,000 of 6,000 | 5,759 of 6,000 |
| Viewers who missed frames | 0 | 0 |
| Median delivery | 61 ms | 16–20 ms |
| p95 delivery | 134–136 ms | 25–31 ms |
| Slowest frame | 0.38 s | 0.17–0.69 s |
| Delivery confirmed by the relay's own counters | Yes | No, none exposed |
| Infrastructure | One 4-vCPU VM running MOQOM's relay, at 97% CPU | Cloudflare's global network (beta) |
What Cloudflare's beta relay leaves out
| Feature | Cloudflare MoQ (beta) | moqom.cloud and self-hosted MOQOM |
|---|---|---|
| Post-quantum end-to-end encryption (SFrame, MLS room keys) | Yes, keys from MOQOM | Yes |
| Protocol version | draft-16 | draft-22, and draft-19 |
| Catch-up and late join from cache (FETCH) | No | Yes |
| Pause, resume and priority changes (REQUEST_UPDATE) | No | Yes |
| Graceful relay migration (GOAWAY) | No | Yes |
| Per-viewer tokens | No, 10 tokens per relay, shared | Yes, device-bound |
| Token kept out of logs | No, it travels in the URL path | Yes, in-band |
| Admission tickets, kick and ban, spend limits | No | Yes |
| Relay counters, metering, events, recording | No | Yes |
| Relay-side layer and bitrate adaptation | No | Yes |
| Direct (peer-to-peer) calls | No | Yes |
Cloudflare run on 22 October 2026 with the same traffic, viewers opening 100 sessions a second from three Google Cloud load generators in us-west2, frames counted once all were in. Its relay exposes no counters, so its figures are what viewers reported. Delivery times, not glass to glass; the rigs differ, so read the comparison as indicative. MOQOM's SDK runs on Cloudflare's relays too: how.Full test data and developer notes →
54 µs to encrypt a 1080p keyframe end to end
SFrame with AES-256-GCM, on the device, against a 33 ms frame budget. Encrypting media payloads costs a small fraction of a frame.
Measured in lab: 6,000 viewers on one 4-vCPU relay, every frame deliveredHow we measure
- Latency is timestamped in software at capture and read at render, against a shared clock
- A run is marked untrustworthy if the clock error exceeds the median it reports
- Phone-side numbers are checked against the relay's counters. A smooth run that only looked smooth because frames were dropped doesn't count
- Fan-out uses MOQOM's own load generator, which can drive any relay that speaks draft-19, and Cloudflare's on draft-16
- Our relay restarts between sizes, so each size starts from cold state
A physical photodiode rig is planned, and its numbers are the ones we will quote in a contract. Cellular, cross-city, Web, Android and desktop numbers will be published with each release.
No other MoQ stack publishes a glass-to-glass figure measured comparably, so the Cloudflare comparison above is of delivery times only. The fair glass-to-glass test is the same two phones through each relay, and it is on our list.
Test coverage
- Swift SDK
- ~864 tests
- Media engine
- ~944 tests
Plus coverage-guided fuzzing of every decoder on every change, and cross-language interop: a Swift publisher and subscriber through the Rust relay, and both backend SDKs against the real control plane.