Alibaba releases Qwen3.5-Omni with 10h audio input and video support

• Alibaba's launch post introduced Qwen3.5-Omni as a multimodal family for native text, image, audio, and video understanding, with Plus, Flash, and Light variants and a separate realtime SKU also reflected in TestingCatalog's summary.
• The main implementation change is input scale: according to the launch post, Qwen3.5-Omni can handle up to 10 hours of audio or 400 seconds of 720p video, while Alibaba says it also adds script-level captioning with timestamps, scene cuts, and speaker mapping.
• Qwen is positioning the model as both offline and live infrastructure: Alibaba's announcement highlights built-in web search, complex function calling, turn-taking, and voice controls, and the demos in the vibe-coding post and the interruption demo show those features in active use.
• One caveat from early readers is scope: as kimmonismus notes, “omni” here is mostly about interpreting multimodal input rather than full image generation, even though the system does synthesize speech and Alibaba says voice cloning is coming.

Qwen3.5-Omni is a new family rather than a single endpoint. Alibaba's launch post names Plus, Flash, and Light tiers, while the offline API docs and realtime docs split the product into batch-style multimodal processing and a WebSocket-based live interaction stack. The realtime docs say developers can integrate either directly over WebSocket or through the DashScope SDK in Python or Java, with sessions running up to 120 minutes and region-specific endpoints for Beijing and Singapore.

On published evals, Alibaba claims Qwen3.5-Omni-Plus “outperform[s] Gemini-3.1 Pro in audio” and matches it in broader audio-visual understanding launch post. The benchmark table in

backs up a mixed but competitive picture: Qwen leads Gemini 3.1 Pro on DailyOmni, QualcommInteractive, Omni-Cloze, VoiceBench, RUL-MuchoMusic, MMAU, and several vision tasks, while trailing on WorldSense and some text benchmarks.

Alibaba's headline workflow is “Audio-Visual Vibe Coding”: in the demo post, a spoken prompt and camera input produce working code for a site or game, and a second example in another demo shows the same pattern landing a complete website from live multimodal input. That matters less as a flashy demo than as a signal that Qwen is treating camera frames, speech, and code generation as one loop rather than separate model calls.

The other concrete offline workflow is long-form media parsing. In the captioning demo, Qwen generates dense audio-visual captions, and Alibaba says those outputs can be elevated to “script-level captioning” with timestamps, scene boundaries, and speaker mapping launch post. A separate travel-planning demo also shows the model grounding on UI, dates, and itinerary details from mixed visual and spoken input instead of a single uploaded image or clip.

The realtime demos are more specific than the announcement language. In the voice-control demo, the model's speech is adjusted for style, emotion, and volume mid-conversation; in the interruption demo, Alibaba shows “Multi-Turn Dialogue” and “Intelligent Interruption,” matching the launch claim that Qwen can do “smart turn-taking” and ignore noise launch post.

For engineers, the key point is that Alibaba has published both realtime docs and offline API docs at launch. The offline docs describe Python and Node.js quickstarts and note that responses are streaming-only, while the realtime docs add live audio/video ingestion, function calling, and web search. The feature boundary is still worth stating clearly: as one early reaction put it, “omni” does not mean a general image generator, so the release is best understood as a multimodal understanding-and-interaction stack with speech output, not an everything model.