Qwen3.5-Omni Is Here: Alibaba's Omnimodal AI Beats Gemini on Audio

TL;DR

Alibaba released Qwen3.5-Omni on March 30, 2026. It processes text, images, audio, and video in a single model and outputs both text and real-time speech. It outperforms Gemini 3.1 Pro on general audio understanding and reasoning benchmarks, supports 113 languages for speech recognition, and includes voice cloning. Three variants are available: Plus, Flash, and Light.

Try Apidog today

Why Qwen3.5-Omni matters for developers

Most multimodal AI apps are still built as pipelines:

1. Speech-to-text for audio input
2. Vision model for images or video frames
3. LLM for reasoning and response generation
4. Text-to-speech for voice output

That architecture works, but every handoff adds latency, cost, and failure points.

Qwen3.5-Omni collapses that stack into one model. It accepts text, images, audio, and video as input, then returns text or speech as output from a single inference call.

The model supports a 256,000-token context window, which can fit:

• More than 10 hours of audio
• Roughly 400 seconds of 720p video with audio
• Around 190,000 words of text

Alibaba trained it on more than 100 million hours of native audio-visual data. The practical result: you can build applications where the model reasons across speech, visuals, and text together instead of stitching together separate model outputs.

Use cases include:

• Voice assistants
• Video analysis tools
• Multilingual support agents
• Accessibility tools
• Developer productivity workflows using screen recordings

What changed from Qwen3-Omni

The previous generation, Qwen3-Omni Flash, launched in December 2025 with 234ms response latency. Qwen3.5-Omni is the next full release.

1. Speech recognition now covers 113 languages

Qwen3-Omni supported speech recognition in 19 languages. Qwen3.5-Omni expands that to 113 languages and dialects.

Speech generation also increased from 10 languages to 36.

For developers, this matters if your app needs to support users outside major Western markets. Instead of routing different languages through separate ASR vendors, you can test one model across a much broader language set.

Example workflows:

• Transcribe customer support calls in multiple languages
• Summarize non-English podcasts or interviews
• Handle bilingual conversations with mid-sentence language switching

2. Voice cloning is available through the API

Qwen3.5-Omni Plus and Flash support voice cloning through the API.

The workflow is:

1. Upload or reference a voice sample
2. Send the user prompt and voice configuration
3. Receive speech output in the cloned voice

This is useful for:

• Consistent voice personas
• Branded voice agents
• Long-running conversational assistants

Voice cloning was not available in the previous generation.

3. ARIA improves pronunciation for numbers and technical terms

Neural TTS systems often struggle with:

• Product names
• Technical terms
• Proper nouns
• Prices
• Version numbers
• Acronyms

Qwen3.5-Omni introduces ARIA, a dynamic text-speech synchronization layer. It reads ahead in the text buffer and adjusts phoneme generation before audio is emitted.

That helps terms like these render correctly:

IPv6
$249.99
Qwen3.5-Omni

This is especially relevant for developer tools, sales assistants, support bots, and technical documentation readers.

4. Semantic interruption improves voice UX

In many voice systems, any incoming audio is treated as an interruption.

That creates bad UX:

• User says “uh-huh” → assistant stops
• User says “right” → assistant stops
• User says “wait, stop” → assistant should stop

Qwen3.5-Omni distinguishes between backchannels and real interruptions.

This makes it more useful for real-time voice assistants where users naturally acknowledge, pause, or interrupt during conversation.

5. Real-time web search is integrated

Qwen3.5-Omni can query the web during inference and include live results in its response.

That means you do not always need to pre-fetch external context and inject it into the prompt yourself.

Use this when your app needs current information, such as:

• Recent documentation
• Current pricing
• News or market updates
• Live product information

6. Screen recordings can become coding input

Qwen3.5-Omni supports “Audio-Visual Vibe Coding.”

The workflow:

1. Record a screen interaction
2. Send the video to the model
3. Ask the model to replicate, explain, or improve what it sees
4. Use the generated code as a starting point

This gives coding assistants a new input format: video.

Instead of describing UI behavior in text, you can provide a screen recording and ask for implementation guidance.

Benchmark results

Across 36 audio and audio-visual benchmarks:

• Qwen3.5-Omni achieves state-of-the-art on 32 out of 36
• It sets new state-of-the-art on 22 of those 36
• It outperforms Gemini 3.1 Pro on general audio understanding, reasoning, and translation
• It matches Gemini 3.1 Pro on audio-visual comprehension

For speech generation quality, Alibaba reports that Qwen3.5-Omni beats ElevenLabs, GPT-Audio, and Minimax on multilingual voice stability across 20 languages.

That comparison is notable because ElevenLabs is a dedicated voice AI company. Still, you should benchmark against your own prompts, languages, accents, and audio conditions before choosing a production model.

Model variants

Alibaba ships three versions.

Variant	Best for
Qwen3.5-Omni Plus	Maximum quality; audio-visual reasoning, voice cloning, long-context tasks
Qwen3.5-Omni Flash	Balanced speed and quality; real-time voice chat, production APIs
Qwen3.5-Omni Light	Low-latency tasks; mobile and edge scenarios

All three handle text, images, audio, and video as input.

The differences are mainly:

• Output quality
• Latency
• Cost
• Deployment fit

For most production APIs, start by testing Flash. Use Plus when quality matters more than latency or cost. Use Light for latency-sensitive scenarios.

Working with the 256K token context window

Qwen3.5-Omni supports up to 256,000 input tokens.

In practice, that means you can send:

• A long meeting recording
• A full support call
• A product demo video
• A large document
• Mixed text, audio, image, and video context

Approximate capacity:

Input type	Approximate capacity
Audio	More than 10 hours of continuous speech
Video	Roughly 400 seconds of 720p video with embedded audio
Text	Around 190,000 words

For many multimodal use cases, this removes the need to chunk inputs manually.

Example prompt for a meeting recording:

Summarize this meeting recording.

Return:
1. Main decisions
2. Open questions
3. Action items with owners
4. Risks mentioned
5. Follow-up email draft

Example prompt for a product demo video:

Analyze this product demo video.

Return:
1. What feature is being demonstrated
2. Step-by-step user flow
3. Any visible bugs or UX friction
4. Suggested improvements
5. A short release note

Qwen3.5-Omni’s 256K context is smaller than Gemini 2.5 Pro’s 1M context and larger than many standard multimodal workflows require. Compared with GPT-4o’s 128K context, it gives more room for long audio-visual inputs.

Building multilingual voice workflows

The increase from 19 to 113 speech recognition languages changes how you can design multilingual systems.

Customer support

You can route voice input from many regions into the same model rather than maintaining separate ASR pipelines per language.

Example support-agent prompt:

You are a customer support assistant.

Input:
- Customer audio
- Product documentation
- Recent order information

Tasks:
1. Detect the spoken language
2. Transcribe the user request
3. Identify the issue
4. Reply in the same language
5. Escalate if the issue involves billing, legal, or account security

Content processing

For podcasts, interviews, and videos, one request can cover transcription, translation, and summarization.

Example:

Process this interview.

Return:
1. Original-language transcript
2. English translation
3. 5-bullet summary
4. Notable quotes
5. Speaker-by-speaker topic breakdown

Mid-conversation language switching

Bilingual users often switch languages mid-sentence.

Qwen3.5-Omni handles this natively, which is useful for:

• Support conversations
• Education apps
• Travel assistants
• Internal company tools for global teams

Architecture: Thinker-Talker with MoE

Qwen3.5-Omni uses a Thinker-Talker architecture.

The Thinker processes multimodal input and generates reasoning tokens.

The Talker converts those tokens into natural speech in real time using a multi-codebook approach designed to reduce latency.

The Plus variant uses Mixture of Experts, or MoE. With MoE, only a subset of model parameters activates per token. This keeps inference more efficient than a dense model of equivalent quality.

For local deployment:

• Use vLLM when possible for MoE-optimized serving
• Use HuggingFace Transformers if you need broader compatibility
• Expect higher latency with Transformers on MoE architectures

Basic local deployment decision flow:

Need production local serving?
→ Start with vLLM

Need experimentation or custom model loading?
→ Try HuggingFace Transformers

Limited GPU memory?
→ Test Flash or Light before Plus

Testing Qwen3.5-Omni APIs with Apidog

If you evaluate Qwen3.5-Omni through an API, you will likely send multimodal request bodies containing:

• Text prompts
• Image URLs
• Base64-encoded audio
• Video references
• Model variant settings
• Streaming options
• Voice cloning parameters

This gets hard to debug with raw curl commands.

Apidog is useful for building, saving, and testing these request templates. You can:

• Store DashScope API keys as environment variables
• Create reusable request bodies for Plus, Flash, and Light
• Compare latency across variants
• Validate response structure
• Write automated tests for expected fields

Example API test checklist:

[ ] Request returns HTTP 200
[ ] Response includes text output
[ ] Response includes audio output when requested
[ ] Latency is within target range
[ ] Language detection matches input
[ ] Streaming response starts before full generation completes
[ ] Error responses are handled correctly

A typical multimodal request template might include:

{
  "model": "qwen3.5-omni-flash",
  "input": {
    "messages": [
      {
        "role": "user",
        "content": [
          {
            "type": "text",
            "text": "Summarize this video and extract action items."
          },
          {
            "type": "video_url",
            "video_url": "https://example.com/demo.mp4"
          }
        ]
      }
    ]
  },
  "parameters": {
    "output_modalities": ["text", "audio"],
    "language": "auto"
  }
}

Use the same saved request against Plus, Flash, and Light to compare output quality and latency under identical inputs.

Download Apidog free to start testing multimodal API requests.

Who should evaluate Qwen3.5-Omni?

Qwen3.5-Omni is worth testing if you are building any of the following.

Voice assistants

Use it for real-time speech input and speech output with conversation memory and web retrieval.

The semantic interruption and ARIA features address two common voice UX problems:

• False interruptions
• Mispronounced technical terms

Video analysis tools

Use it for:

• Meeting transcription
• Product demo analysis
• Tutorial generation
• Video summarization
• Screen-recording-to-code workflows

The 256K context window means many recordings can fit in one request.

Multilingual customer products

Use it when your app needs:

• 113-language ASR
• 36-language TTS
• Language switching
• One multimodal model instead of multiple vendors

Accessibility tooling

Potential workflows include:

• Alt text generation
• Audio descriptions for video
• Real-time captions
• Multilingual caption generation

Developer productivity tools

Audio-Visual Vibe Coding lets developers provide screen recordings as context.

Example prompt:

Watch this screen recording of a UI interaction.

Generate:
1. React component structure
2. Required state management
3. CSS layout
4. Edge cases
5. A minimal working implementation

Access options

Qwen3.5-Omni is available through:

• Alibaba Cloud DashScope API for production API access
• qwen.ai for web-based testing
• HuggingFace Hub for model weights and local deployment
• ModelScope, recommended for users in mainland China

The API follows Alibaba Cloud’s standard authentication model. You need a DashScope API key.

Check the DashScope documentation for:

• Endpoint details
• Authentication
• Streaming support
• Pricing by modality
• Rate limits
• Model availability

What to test before production

Benchmarks are useful, but your own workload matters more.

Before adopting Qwen3.5-Omni, test:

• Your users’ accents
• Your supported languages
• Domain-specific vocabulary
• Audio quality from real devices
• Long recordings
• Noisy environments
• Video formats and resolutions
• Latency under load
• Voice cloning quality
• Streaming behavior

Also note:

• Voice cloning is API-only for now
• The qwen.ai web interface does not expose voice cloning yet
• Local deployment requires significant GPU memory
• The Plus variant, a 30B MoE model, needs at least 40GB VRAM for comfortable inference
• Flash and Light are more accessible for smaller deployments

FAQ

How is Qwen3.5-Omni different from Qwen2.5-Omni?

Qwen2.5-Omni supported 7B and 3B dense model sizes with 19 languages for speech. Qwen3.5-Omni uses an MoE architecture, expands speech recognition to 113 languages, adds voice cloning, and introduces ARIA for better audio quality. Benchmark performance and context length also increased.

Can I run Qwen3.5-Omni locally?

Yes. You can run it with HuggingFace Transformers or vLLM.

For production local deployment, vLLM is the better option because it handles MoE routing more efficiently.

The Plus variant needs 40GB+ VRAM. Flash and Light run on smaller GPUs.

Is there a free tier?

The qwen.ai web interface is free to use. API access through DashScope is paid.

Pricing depends on modality, such as audio tokens, video frames, and text tokens. Check the DashScope pricing documentation for current details.

Does it support real-time streaming?

Yes. The Thinker-Talker architecture outputs audio in streaming chunks, so the first audio bytes can arrive before the full response is generated.

This is important for live voice conversations.

What is the difference between Plus, Flash, and Light?

Plus is the highest-quality variant and is best when accuracy matters more than speed.

Flash balances speed and quality and is the best starting point for most production APIs.

Light is the fastest option and is intended for latency-sensitive applications such as mobile or edge scenarios.

Can I use my own voice with the API?

Yes. Voice cloning is available through the API.

You provide an audio sample of the target voice, and the model uses it for speech output.

This feature is not available through the web interface yet.

How does it compare to ElevenLabs for voice generation?

On Alibaba’s benchmarks across 20 languages, Qwen3.5-Omni Plus outperforms ElevenLabs on multilingual voice stability.

ElevenLabs still has a longer product track record and more voice-specific customization options. If you only need voice generation, compare both. If you need one integrated multimodal model, Qwen3.5-Omni is the cleaner architecture to test.

Is it safe to send sensitive audio or video data through the API?

Review Alibaba Cloud’s data processing agreement before sending sensitive data.

As with any cloud API, assume data may be logged unless the agreement explicitly states otherwise.