Alibaba / Qwen · 2026-03-30 · major
Qwen 3.5-Omni — native omnimodal model
Native omnimodal model processing text, images, audio, and video in a single pass with real-time speech I/O, voice cloning, and 113-language speech recognition. Outperforms Gemini 3.1 Pro on audio tasks.
A single model that natively sees, hears, reads, and speaks — with 113-language speech recognition and real-time voice cloning from a 10-second sample.
Key specs
| Context window | 256K tokens |
|---|---|
| Speech recognition languages | 113 |
| Speech generation languages | 36 |
| Sota results (plus) | 215 |
What is it?
Qwen3.5-Omni is Alibaba's latest omnimodal model, released March 30, 2026. It processes text, images, audio, and video natively in one forward pass and generates both text and streaming speech output in real time. It ships in three variants: Plus (performance leader), Flash (recommended for production), and Light (edge and budget use). The Plus and Flash variants are API-only; the Light variant is open-weights on Hugging Face.
How does it work?
The architecture uses a Thinker-Talker design with Hybrid-Attention Mixture of Experts across all modalities. A standout feature is native turn-taking intent recognition, which distinguishes between background noise and genuine user interruptions, enabling full-duplex conversations. Voice cloning requires only 10-30 seconds of sample audio. The model handles 256K token context, 10+ hours of audio input, and 400+ seconds of 720p video at 1 FPS.
Why does it matter?
This is the most complete omnimodal package from any Chinese lab, directly competing with Gemini 3.1 Pro on multimodal benchmarks. The Plus variant set 215 SOTA results across audio and audio-video benchmarks. For teams building voice-first or multimodal agents, the breadth of modalities in one model eliminates the need to stitch together separate speech, vision, and language models.
Who is it for?
Developers building multimodal or voice-first AI products, teams evaluating alternatives to Gemini for omnimodal tasks.
Try it
huggingface.co/spaces/Qwen/Qwen3.5-Omni-Offline-Demo