Alibaba releases Qwen3.5-Omni, a native multimodal model for text, audio and video

Alibaba has shown Qwen3.5-Omni, a native omnimodal model that understands text, images, audio and video in a single architecture and can respond by voice in real time.

How the model is built

The core idea behind Qwen3.5-Omni is that it is not a set of separate models bolted onto a text core, but a single system designed from the start for several types of data. Alibaba contrasts this approach with the older style of multimodality, where vision or sound was simply "sewn onto" an LLM through external encoders. For developers the difference matters: a native architecture usually holds context better across channels, links speech to images more accurately and scales more easily to real-world scenarios such as calls, video analytics and voice assistants.

The technical report describes Qwen3.5-Omni as a member of the omnimodal family with a Hybrid Attention Mixture-of-Experts architecture for two components, Thinker and Talker. The first handles understanding and reasoning; the second produces the streaming voice response. Qwen says the model was trained on heterogeneous text-image pairs and on more than 100 million hours of audiovisual data.

The stated context window is 256 thousand tokens, which means a single session can include very long conversations, meeting recordings, lectures, screenshots and video clips without splitting them into dozens of small requests.

The series comes in several variants: Plus, Flash and Light. This hints at a familiar product-line logic: maximum quality for hard tasks, a fast mode for interactive scenarios and a lighter version to save compute.

Alibaba also emphasizes real-time operation: Qwen3.5-Omni can stream its answer as text and as natural speech, and a mechanism called ARIA, which dynamically aligns text and speech units, is responsible for more stable and smoother voice generation.

Key capabilities of the release

According to the technical report, Qwen3.5-Omni-Plus achieves the best results on 215 tasks and benchmarks covering audio and audiovisual understanding, reasoning and interaction. Qwen specifically highlights that the model outperforms Gemini 3.1 Pro on key audio tasks and is on a comparable level in comprehensive audiovisual understanding.

For Alibaba this is an important signal to the market: competition among top multimodal models is no longer limited to OpenAI and Google, and Chinese labs are laying claim to leadership in the hardest modes, namely voice, video and live dialogue.

• 256k context window
• More than 10 hours of audio in one session
• More than 400 seconds of 720p video at 1 FPS
• Plus, Flash and Light variants
• Structured captions with scenes and timings

Another strong part of the release is audio and video description. The report mentions script-level structured captions: the model can build detailed descriptions with precise time synchronization and automatic scene segmentation. This is useful not only for media archives but also for video database search, call analytics, education, accessibility scenarios and content quality control. In effect, Alibaba is pushing Qwen3.5-Omni toward a universal understanding layer for any media format, rather than just a "chatbot that can also hear."

The researchers also note a new capability they call Audio-Visual Vibe Coding. This means coding directly from audiovisual instructions: the model can interpret not only a text prompt but also a spoken explanation together with visual context. For now it is more a research signal than a ready mass-market product, but the direction is telling. If such modes take hold, a developer could skip retyping a bug report by hand, and instead show the interface, explain the problem aloud and get a working draft of a solution.

What this means

Qwen3.5-Omni shows that the next stage of the AI race is not about yet another text chatbot, but about models that work equally confidently with sound, images, video and speech in a single stream. For business, this opens the way to more cohesive products: voice agents, meeting analysis, media search and interfaces that understand not only text but everything the user shows and says.