Qwen 3.5 Omni: Alibaba’s AI Model Can Now Hear, Watch, and Clone Your Voice

Alibaba just dropped its most ambitious AI upgrade yet.

The company’s Qwen team released Qwen 3.5 Omni on Sunday, a new version of its “omnimodal” AI that simultaneously processes text, images, audio, and video, and talks back in real time across 36 languages, placing its model on the same battlefield as the latest state-of-the-art AI foundational models currently available.

“Omni” isn’t just a marketing buzzword here. Most AI models you interact with are primarily text-in, text-out systems. Some handle images, some handle voice. Qwen 3.5 Omni handles all of them natively, at the same time, without the need to convert everything to text through third-party tools.

The new model comes in three sizes—Plus, Flash, and Light—all supporting a small (by today’s standards) 256,000-token context window. It was trained on over 100 million hours of audio-visual data—a scale that puts it in a different weight class from most competitors.

Qwen 3.5 Omni is an evolution of Qwen 3 Omni Flash, Alibaba’s previous omnimodal model released in December 2025. That version already impressed with its ability to process video and audio simultaneously—it could handle image editing instructions combining multiple visual inputs in ways competitors couldn’t—and streamed voice responses with latency as low as 234 milliseconds.

It was also the first model to try an alternative to Google’s NotebookLM. It achieved something, but the quality was not on par with Google’s offer.

Qwen 3.5 Omni takes all of that and adds a longer context window, better reasoning, a much wider language library, and a set of real-time interaction features the previous generation didn’t have.

The headline upgrade is what happens when you actually talk to it. Qwen3.5-Omni now supports semantic interruption: It can tell the difference between you saying “uh-huh” mid-sentence and actually wanting to cut in, so it won’t stop mid-thought every time someone coughs in the background, making spoken interaction more seamless.

A new technique called ARIA, short for Adaptive Rate Interleave Alignment, also fixes a subtle but persistent annoyance: AI systems that garble numbers or unusual words when reading aloud. ARIA dynamically syncs text and speech to keep output natural and accurate.

Then there’s voice cloning. Users can upload a voice sample and have the model adopt that voice in its responses, a feature that puts Qwen directly in competition with ElevenLabs and other dedicated voice tools. We weren’t able to access this feature, though, because this is a feature that, at least for now, is only available via API..

On multilingual voice stability benchmarks, Qwen3.5 Omni- Plus beat ElevenLabs, GPT-Audio, and Minimax across 20 languages. The model also now supports real-time web search, meaning it can answer questions about breaking news or live market data without pretending it already knows.

The team is also highlighting what they’re calling “Audio-Visual Vibe Coding,” the model can watch a screen recording or video of a coding task and write functional code based purely on what it sees and hears, no text prompt required. It’s a small preview of how AI assistants might eventually operate inside your workflow rather than alongside it.

To understand what “omnimodal” actually means in practice, we ran a quick test: We fed both Qwen3.5-Omni and ChatGPT 5.4 in “thinking” mode the same YouTube Short—a clip of Dastan President (Dastan is Decrypt’s parent company) and commentator Farokh discussing breaking news. Qwen 3.5 Omni processed the video natively and returned a full analysis in about one minute: who was speaking, what they were discussing, and a substantive comment on the topic based on its own knowledge of the subject area.

ChatGPT 5.4, which is not omnimodal, had to manage with what it got. It extracted frames from the video, ran them through a vision model, used Whisper to transcribe the audio, and applied an OCR tool to read embedded subtitles—three separate processes stitched together to approximate what Qwen3.5-Omni does in a single pass. The result took nine minutes, and that’s under ideal conditions: a well-lit video with clean audio and burned-in subtitles. Real-world content rarely offers all three.

In our quick tests across multiple inputs, the model also handled prompts in Spanish, Portuguese, and English without issue—switching languages mid-conversation without losing context.

On standard benchmarks, Qwen 3.5 Omni Plus outperformed Gemini 3.1 Pro on general audio understanding, reasoning, and translation tasks, and matched it on audio-visual comprehension. Speech recognition now covers 113 languages and dialects—up from 19 in the previous generation.

This is Alibaba’s second major AI release in six weeks. In February, it launched Qwen 3.5, a text-and-vision model that matched or beat frontier models on reasoning and coding benchmarks—part of a streak that has also included Qwen Deep Research and a lineup of tools rivaling OpenAI and Google. Qwen 3.5 Omni extends that momentum into full multimodal territory, at a time when every major AI lab is racing to build systems that handle the full spectrum of human communication—not just words on a screen.

The model is available now via Alibaba Cloud’s API and can be tested directly at Qwen Chat or through Hugging Face’s online demo.