Muse Wonder reviews a video: send Sume inspect stills and transcript

Muse Wonder takes video and audio in but replies only in text. To have it review a Sume render, use video_inspect: 8 stills by default, 24 at most.

4 min readSume
All posts

To have Meta's Muse Wonder review a clip you rendered on Sume, call video_inspect and pass the model its stills and transcript as text and images. By default Sume returns 8 stills at up to 768 px, and a single call can return at most 24 stills from a source up to 1,800 seconds long.

Meta's model page lists text, image, video, audio and PDF as input, text as the only output, and a 1,048,576-token context window (read 2026-10-13).

Why not hand the model the MP4 directly?

Native video input may work, but the sources I read give no limits for it: no maximum duration, file size or frame rate. Sume's route has documented limits and returns small, checkable pieces. The video inspect page says the route reads one media.sume.com clip the workspace already owns, never re-encodes it and never makes an MP4.

A clip from another host has to be imported first with POST /v1/media-imports; the API rejects off-host URLs at admission. Inspect then returns a probe object, frames[{t,url,width,height}] as durable image artifacts, and optionally a transcript.

What does one inspect call give the model?

The frames field controls the stills. Only one of at[] or fps may be sent, never both. Raise max_edge (64 to 2160) if the reviewer must read on-screen text. Use seek: "precise", the default, when the timestamp matters; fast can return a still up to one GOP early, roughly 0 to 5 seconds on typical sources.

video_inspect stills options, from Sume docs (read 2026-10-13)
frames valueWhat comes backLimit
not sent8 mid-bin stills (1 fps if the clip is under 8 s)768 px longest edge
falseProbe only, no stillsnone
{ at: [seconds] }Stills at the instants you name1 to 24 values, each 0 or more
{ fps: n }Evenly sampled stills0 < n <= 2, capped at 24 stills

How do I give it the audio?

Send transcribe: true and the response includes transcript with text, words[] and optional sentence segments[]. This is the billed part: Sume reserves one minute unless you send duration_seconds (maximum hint 600), and the credits scale with audio minutes at the API rate card.

Check probe.has_audio first with a frames: false call. A silent clip with transcribe: true returns inspect_source_has_no_audio.

How does the agent wire it together?

Inspect is the hosted MCP tool video_inspect. A write needs idempotency_key (and mcp:write under OAuth). The default mode is sync and waits at most 30 seconds; on a 202 the agent polls with jobs_wait and then jobs_result, as listed on MCP tools and gates.

Muse Wonder is listed on OpenRouter as a chat-completions model, so the review call is a normal chat request that carries the still URLs and transcript text. Whatever it replies is text; to fix what it finds, the agent calls Sume's generation tools again. Keep the stills small: eight 768 px frames and a transcript are enough to catch a wrong product, a garbled caption or a mismatched voice, and each still is a durable media.sume.com URL the model can cite back by timestamp t.

  • Review a 30-second ad: fps: 0.5 gives 15 stills, as in this walkthrough.
  • Check a long render: 24 stills plus a transcript, as in this one.
  • Start from a URL: import first, then inspect, as in this one.

Sources

Related posts

More in Models

All Models posts

Written by Sume