Choose a model for multimodal understanding, audio and video analysis, voice conversation, content moderation, or speech translation.
Omni-modal models understand text, audio, images, and video. Qwen3.8-Omni-Flash supports audio/video analysis, meeting summaries, and subtitle generation, with thinking mode, tool calling, and web search. Three series are available:
For dated versions, the capability matrix, and translation language coverage, see the Recommended models, All models, and Translation sections of Speech-to-speech models. For the voices each model supports, see Voice list.
| Series | Position | Highlights |
|---|---|---|
| Qwen3.5-Omni | Flagship | Most capable. Web search and voice cloning. Up to 3 hours of audio or 1 hour of video |
| Qwen3-Omni-Flash | Lightweight | Lower cost, thinking mode. 150 s per request |
| Qwen3.5-Livetranslate | Translation | Works out of the box. 60 languages, ~3 s latency |
Choose a model by scenario
| Scenario | Recommended model | API | Guide |
|---|---|---|---|
| Audio/video understanding and text generation — analyze audio/video content and generate meeting summaries, subtitles, and text answers | Qwen3.8-Omni-Flash | Chat Completions / Responses | Audio and video file understanding |
| Real-time voice and video chat — interact through microphone and camera (voice assistants, customer service, visual Q&A, live-stream analysis) | Qwen3.5-Omni Realtime | WebSocket | Realtime audio and video understanding |
| Real-time voice chat with semantic VAD — end-to-end voice interaction where filler sounds do not interrupt; supports function calling | Qwen-Audio Realtime | WebSocket | Realtime audio chat |
| Offline audio output — upload audio or video files and generate speech responses | Qwen3.5-Omni | Chat Completions | Audio and video file understanding |
| Real-time speech translation — simultaneous interpretation, ~2.3 s latency, 60 languages (conferences, live translation) | Qwen3.5-Livetranslate | WebSocket | Real-time audio and video translation |
| Audio and video file translation — translate an uploaded file into a target language (video dubbing, podcast translation) | Qwen3-Livetranslate | Chat Completions | Audio and video file translation |
| Voice cloning — supply reference audio and reply in that voice | Qwen3.5-Omni Plus / Flash | Chat Completions / Realtime API | Voice cloning |
Capability boundaries
- Function calling — Qwen3.8-Omni-Flash (Chat Completions / Responses), Qwen3.5-Omni Plus / Flash (HTTP, text output), Qwen3-Omni-Flash (HTTP), and Qwen-Audio Realtime (WebSocket).
- Web search — Qwen3.8-Omni-Flash (Chat Completions / Responses) and Qwen3.5-Omni (Chat Completions / Realtime API). Qwen-Audio Realtime (3.0 Plus/Flash, 3.1 Plus) also supports web search via
enable_search; cannot be enabled at the same time as Function Calling. Qwen3.5-Omni cannot enable web search and function calling at the same time. - Thinking mode — Qwen3-Omni-Flash (HTTP) only, and it returns no speech output.
- Legacy models —
qwen-omni-turbosupports Chinese and English only and is no longer updated. Use Qwen3.5-Omni for new projects.
Migrate from closed-source models
| Closed-source examples | QwenCloud recommendation | |
|---|---|---|
| Audio/video understanding and text generation | Gemini 3.8 Flash | qwen3.8-omni-flash |
| Real-time translation | Gemini 3.5 Live Translate | qwen3.8-livetranslate-flash-realtime |
For speech output, use qwen3.5-omni-plus.
Full model list
For dated versions, the capability matrix, and translation language coverage, see the Recommended models, All models, and Translation sections of Speech-to-speech models. For the voices each model supports, see Voice list.
If you only need speech in and speech out, and you are choosing between a single S2S model and a pipeline (ASR + LLM + TTS), start with Speech-to-speech models.