• Meta's September 23, 2026 research post reports about 870 ms from the end of a user's turn to the first byte of a voice-and-video response, on a stated configuration.
• The 12-concurrent-sessions and 8x serving-capacity figures are for a single GB200 with Meta's own optimization stack, and no other hardware is reported.
• The 78%-to-22% and 88%-to-12% rater preferences over Runway Characters and HeyGen LiveAvatar come from Meta's own tests, with no rater count or method given.
• The Verge reports a real-time voice mode reaching the Muse iOS app, but neither source says it runs on the models Meta benchmarked.
• Verdict: the engineering claims are specific, but they are unreproduced vendor claims, and the shipped product is untested by outsiders.

Meta's Muse agent now supports real-time voice conversation, and the only hard numbers behind Meta's voice-and-avatar stack come from Meta itself. Its September 23, 2026 research post reports roughly 870 ms response latency and 12 concurrent sessions per GB200. Both were measured by the vendor on one configuration. Nothing we could find ties those figures to the voice mode The Verge describes reaching users, so they are leads to test, not proof.

What Meta's own post claims

Meta's research post describes Muse Realtime Voice, which emits speech tokens, and Muse Realtime Avatar, which turns those tokens into video, as "a single streaming system." It reports 448x768 portrait video at 25 frames per second, approximately 870 ms from the end of a user's turn to the first byte of the synchronized response, and 12 concurrent video-generation sessions on a single GB200, an 8x serving-capacity gain over a two-step BF16 baseline.

The latency figure is narrower than it sounds

The 870 ms number is one measurement point, the first byte of the response, under one stated hardware configuration. The post gives no distribution, so a median cannot be told apart from a tail, and no results for other hardware or for phone networks. Meta also states that its examples "illustrate model capability and do not all reflect avatars available in the Muse app." First-byte latency is not the same as a conversation feeling responsive.

The preference tests are Meta's own

Meta says raters held two- to three-minute conversations with Runway Characters and HeyGen LiveAvatar, then preferred Muse Realtime Avatar 78% to 22% and 88% to 12% respectively. The post does not give the number of raters, how they were recruited or the statistical method, though it notes one mannerism comparison was not distinguishable from parity. Without those details a reader cannot reproduce or challenge the margins, so this is a vendor claim, not an independent benchmark.

What the voice-mode report adds

The Verge reports that Mark Zuckerberg announced real-time voice conversations with the Muse agent, that the feature reached the reporter's account and that the reporter tried it in the iOS app. We could read only the published excerpt. Neither it nor Meta's post says the shipped voice mode runs on Muse Realtime Voice or Muse Realtime Avatar, so the benchmark figures should not be attributed to what users hear in the app.

Verdict: specific claims, unproven product

The engineering claims are concrete enough to be tested, and that counts in Meta's favor. They are still unreproduced, measured on Meta's hardware and not linked to the app's voice mode. Teams weighing Muse for customer-facing use should wait for independent latency tests on the shipped product. Casual users face little risk. Meta states that Muse is for users aged 18+ and that Video Seal watermarks generated video, which are stated safeguards, not tested ones.