50 seats total·128 fps median measured over 16 min · Helio 0.1.115 · free invite-only beta.→
Evidence

Dated sessions, not one magic number

The figures below come from logs and counters from real sessions. They describe one setup at one point in time; they are not a guarantee for every network.

Selection from Lume market analysis v0.9, reviewed August 16, 2026.

128.4 fpsmedian · direct · 16:04
11.31 MbpsAV1 1440p · median
11 msmedian RTT · AV1 session
20.2 msRG557 median decode
Runs

Representative sessions

Median and P95 are shown together when available. Very short or incomplete sessions are labelled.

Date and pathSetupMeasured resultHonest reading
08 Jul · AV1 direct2560×1440 · AMD → Chrome11.31 Mbps median · P95 14.74 · 60.1 fps · RTT 11/18 ms · decode 0.9 msGood AV1 path for this pair only; not proof for every browser.
11 Aug · HEVC direct1920×804 · 16:046.59 Mbps median · P95 7.84 · 128.4 fps median · P95 140 · RTT 25/32 msOver 120 fps proven on this run; locked 144 fps not proven.
11 Aug · HEVC direct 8M target1920×804 · 12:547.46 Mbps median · 60 fps · RTT 25/35 ms · 5 freezes / 4.539 sLong usable run, with measured loss and freezes.
28 Jul · Android RG5571920×1080 · HEVC5.10 Mbps median · 31.5 fps · RTT 39.5/48 ms · decode 20.2 msThis small device’s decoder and buffer are the limiting factors.
12 Aug · Windows native localH.264 then HEVC · 5 s each14.18 then 11.46 Mbps · 50.9 then 61.9 fps · decode P95 0.05/0.34 msVery short local proof; WAN and signed distribution not validated.
12 Aug · HEVC direct1920×1080 · 26 min3.17 Mbps median · 59.9 fps · RTT 12/20 ms · decode 1.5/2.1 ms · 27 freezes / 6.975 sStable bitrate and decode, but measured freezes and discontinuous audio.
11 Aug · relay / background92 min1.37 Mbps median · RTT 25.5 ms · buffer P95 478 · 471 freezes / 403.576 sLong incident: this scenario is not acceptable quality.
05 Aug · Apple TV 4K4K60 · tvOS · ~60 s49 freezes / 15.741 s · decode 11–15 ms · audio and ICE activeHost capture/encode incident; proof that resolution alone is not enough.
Protocol

How results are read

  1. Identify the runDate, duration, client, resolution, codec and direct or relay path stay attached to the result.
  2. Read the distributionMedian describes the center of a session; P95 reveals spikes hidden by a simple average.
  3. Separate stagesFrame rate, RTT, decode time, buffer, loss and freezes measure different things.
  4. Publish the caveatA short, local, audio-less or disrupted run keeps that limit explicit.
Definitions

What these counters do not prove

  • WebRTC RTT is not glass-to-glass latency.
  • NAL-to-display or client decode does not cover the full capture, encode and physical display path.
  • A frame-rate peak is not the same as a locked cadence throughout a session.
  • One GPU, phone or network does not represent the whole range.
Glass-to-glass measurement still to be produced

A proper proof requires a high-frame-rate external camera filming the host action and client display together. Until that protocol is run, Lume does not publish a glass-to-glass figure.

Follow new validations

The changelog and roadmap publish fixes, experiments and remaining work.

Open the changelog
Lume © 2026