I Fed a 76-Minute Keynote to Qwen3.8-Omni-Flash: Three Minutes, 14 Sections, One Wrong Verdict

Everyone recaps the keynote. I wanted to query it

You remember a slide, a demo, and have no idea where on the timeline it lives. Hence the "everything in three minutes" recaps: someone else watches, you skim.

Two things landed the same week. On September 15, iOS 27 and macOS 27 shipped, pulling the WWDC26 announcements into their usable phase. On September 17, the Qwen team released Qwen3.8-Omni-Flash, a natively omni-modal model on Bailian, Alibaba Cloud's model platform, with audio input pricing down more than 98%.

I read the coverage: benchmark reposts, demo retellings, nobody who took one genuinely long video and ran it end to end. I had the full WWDC26 recording on disk, 76 minutes with audio intact, so I ran the test I wanted to read.

The result up front: it works, and the economics are almost absurd. A full pass takes around three minutes and costs a third of a cent, and the five screenshot moments it flagged all checked out against pulled frames. It also has a blind spot I did not see coming: asked whether the footage was real, the same model called the official Apple recording an AI-generated fake. That story is further down.

Setup: one sentence, one command

My path is the same every time: install Bailian's official agent skill package into a coding agent, and it learns the bl command family. One command to install, a free API Key and wire it up. The skill ships routing rules, so the CLI only runs when you name Bailian or a specific model.

Every round started from the same sentence:

Using Bailian's qwen3.8-omni-flash: watch this video and tell me what it covers. It's the full recording of the WWDC26 keynote.

The command the agent assembled:

bl omni --model qwen3.8-omni-flash \
  --video wwdc26-keynote-360p.mp4 \
  --message "This is a complete 76-minute keynote recording. Break the whole event into a chronological content outline." \
  --text-only

Three details: --video takes a local path and the CLI handles the upload; --text-only skips reply synthesis into speech; --model must be explicit, or bl omni defaults to the previous-generation qwen3.5-omni-plus.

Snag one: 907 MB breaks the upload path

My lazy first move was to throw the 540p source file straight at it, all 907 MB. It ran for about five minutes, then died: Request timed out (exit 5). My first thought was a too-short timeout, but the retry, same settings, a smaller file, went through. The wall was file size: a 76-minute high-bitrate recording breaks the upload path before analysis starts.

The fix is one local ffmpeg line:

ffmpeg -i wwdc26-keynote-540p.mp4 -vf "scale=640:360" \
  -c:v libx264 -preset fast -crf 30 -c:a aac -b:a 64k wwdc26-keynote-360p.mp4

907 MB becomes 140 MB at 360p, plenty of resolution for understanding what a keynote says. I resubmitted: two minutes and five seconds for the full 76 minutes, one pass. The transcode took about three minutes of local compute, which is to say, free.

The lesson went straight into my agent's rules file:

Before analyzing long video, transcode locally to under 200 MB. For anything above 500 MB, cut the first 10 minutes and test-run that first; if recognition looks right, run the full file.

The outline, the brief, and five frames I could verify

The first ask: a chronological outline. Two minutes and five seconds later, back came 14 sections with minute-level time ranges, running from 00:00 to the closing music film at 74:10, whose featured artist it identified on its own: Erick the Architect. The details held up: the "Golden Gate" naming bit, the yellow microbus on stage.

The timestamps are the point: the outline is an index. Child safety runs from 16:40 to 27:20; the Siri AI material starts at 33:10. That separates this from a compressed recap: the structure survives, so you can go back into it.

Second round, for the "just tell me if it is worth watching" crowd: the 8 most important updates, sorted by impact rather than the keynote's running order. Two and a half minutes: daily-touch first (a rebuilt Siri, upgraded child safety), lower-frequency material last. The rollout details came through too: developer beta that day, public beta the following month, official release in the fall.

Third round, the answer I re-read the most: a 10-section illustrated notebook for someone who has not seen the keynote, ending with the 5 most screenshot-worthy moments, each with an MM:SS timestamp. Two minutes and fifty-one seconds. "Illustrated" is on me: it supplies coordinates, I pull frames. What it picked:

1. 04:47: the yellow hippie microbus, one beat before the Golden Gate name drops.
2. 07:44: the Liquid Glass transparency slider, "Ultra Clear" at one end.
3. 29:37: the Apple Foundation Models and Google Gemini graphics merging into one frame.
4. 40:41: the menu Siri generated for a Brazil vs Morocco viewing party, feijoada and tagine included.
5. 65:55: a Spatial Reframing result, cropped photo edges filled back in by a generative model.

Convincing is not verified. To check, convert each timestamp to seconds (04:47 is 287) and pull the frame:

ffmpeg -ss 287 -i wwdc26-keynote-540p.mp4 -frames:v 1 0447.png

Five checks. The microbus is yellow. The slider sits exactly where it said. Both company names appear side by side. The menu really lists feijoada.

Watching turned into querying: the model supplies coordinates, the frame is the evidence, and the check costs one ffmpeg command. For the exact second, you still pull frames and look with your own eyes.

The surprise: it called the real footage a fake

For the last round I wanted a provenance note, so I added one question: is this an official Apple keynote recording, or something a third party produced or faked?

The verdict was blunt: "This is not a recording of an official Apple keynote, but a third-party concept or parody video, very likely AI-generated or synthesized." Six reasons followed. Three of the most telling: every version number is "27" and macOS is named "Golden Gate," neither "ever announced by Apple"; "Liquid Glass shipped in June 2025," which it used to conclude the video "is set in the future"; and the footage is "too clean," read as a sign of generation. That last one was 360p compression artifacting.

Same footage, same model, one differently framed question, and the conclusion flips. I replied with one line of context: this is Apple's official WWDC26 recording. It reversed on the spot and diagnosed itself: the knowledge cutoff. In its training world, WWDC26 has not happened yet, and everything new it has never heard of gets filed as fabrication.

Its own summary: "The misjudgment is not rooted in the video itself. It comes from checking future content against an outdated map of the world."

That line draws a useful boundary. "Understand what a video says" and "judge whether it is real" are two different paths: the first runs on sight and sound, which was reliable here; the second runs on world knowledge, which has an expiration date. As a watching engine it is excellent; for "is this claim true," verify it elsewhere.

Audio only: everything said, everything seen gone

One more contrast. I extracted the audio track, 71 MB, and ran the same kind of summary. Four minutes and twenty-one seconds. Spoken information came through well; everything visual did not. The microbus, the slider, the model-partnership graphic: none of the verified moments showed up. The output also carries transcription seams, with proper nouns drifting between languages.

That draws the line for real use: audio mode is enough for meetings, interviews, podcasts and lectures; keynotes and product demos, where the visuals teach, go video.

What I keep

Three rules went into my agent's instruction file. Rules you keep in your head get forgotten; rules the agent re-reads before every command get followed.

1. Compress before upload. Above 200 MB, transcode locally first; above 500 MB, test the first 10 minutes before the full pass.
2. Separate "what does it say" from "is it true." Content analysis is reliable; for fact-checking, state your context up front or route through something that can search. World knowledge has a cutoff date.
3. Visual evidence in two steps. Get the timestamp, then pull the frame; the frame is the evidence.

The bill

A bill this small comes down to how the model looks at video. Long video is frame-sampled model-side; the official cap is 2,048 frames, so an hour works out to about one frame per 1.7 seconds. Audio bills by the second, 7 tokens per second per the official docs.

The measured bill: usage delta over six full-pass analyses (outline, brief, notebook, authenticity check, self post-mortem, audio-only) came to ¥0.14 total, about two US cents for the whole round. A single pass: three minutes and about a third of a cent.

To audit your own spending, tell the agent "check my omni-flash usage for the last two days" and it runs:

bl usage stats --model qwen3.8-omni-flash --days 2

Put that next to the official price story: the previous generation charged ¥53 per million tokens for audio input; Omni-Flash charges ¥0.8. A 98% cut usually reads like a press release line. In practice: the entire keynote, watched and summarized, for a third of a cent.

Setup: Node.js 18+, npm install -g bailian-cli, bl auth login. API key here, CLI install guide.

Personal practice log, measured September 2026. Verify costs against your own bl usage stats.