Qwen3.5-Omni claims audio-visual coding, timestamped captions, and one-sample voice cloning

• A launch thread claims Qwen3.5-Omni can do “audio-visual vibe coding,” turning whiteboard brainstorm videos into websites and gameplay recordings into playable code launch demo.
• The same thread says the model also produces script-like video logs with frame-accurate timestamps, scene slicing, and character-plus-audio mapping instead of generic summaries captioning demo.
• In realtime use, the launch materials claim one-sample voice cloning, controls for speed, volume, and emotion, plus interruption handling that distinguishes pauses from background noise realtime demo.
• Linked materials around the release say Qwen3.5-Omni has a 256K context window, can handle 10 hours of audio and 1 hour of video, and is available in chat plus offline and realtime APIs specs post access links.

Qwen3.5-Omni is being pitched as a multimodal production model, not just a chatbot. In the launch thread, Hasan Toor says it can infer code directly from visual and audiovisual input, with demos framed around whiteboard-to-website and gameplay-video-to-code workflows launch demo. The linked release materials also point readers to a browser version on Qwen Chat, a report page, and separate offline and realtime API access points via Qwen Chat, offline API, and realtime API.

The technical claims are broad. According to the thread's specs post, Qwen3.5-Omni has a 256K context window, supports up to 10 hours of audio and 1 hour of video in one pass, recognizes 74 languages, and generates 29. Those numbers matter most for long-form creative review: recorded workshops, edit sessions, reference reels, and extended voice interactions fit the way creators actually work.

The clearest creator-facing workflow change is logging and breakdown. The captioning demo claims Qwen3.5-Omni can output frame-accurate timestamps, slice scenes automatically, map characters to audio, and follow custom instructions, turning raw footage into something closer to an edit log or script draft than a plain summary captioning demo. For filmmakers, documentarians, and social teams, that suggests less manual note-taking between ingest and first cut.

The realtime features push in a different direction: live performance and voice interfaces. The thread claims adjustable emotion, speed, and volume, one-sample voice cloning, web search, tool calling, and “semantic interruption” for more natural turn-taking realtime demo. If those demos hold up outside the launch thread, Qwen3.5-Omni is less interesting as a general assistant than as a multimodal layer for prototyping narrated apps, voiced characters, and hands-free creative tools.