Overview
Qwen3.8-LiveTranslate is a real-time simultaneous interpretation model from Alibaba's Qwen team, announced on 19 September 2026. It listens to live speech, optionally with images, and returns the translation as text and synthesised speech while the speaker is still talking. Alibaba describes it as a multilingual real-time audio and video translation model that understands 60 languages and speaks 29 of them.
The launch post says the model is built on an Interleave architecture that improves faithfulness, fluency and conciseness, and that it reduces average lagging (LAAL) from 2.8 s to 2.3 s across 60 languages. The launch graphic names four additions: synchronized source and translation, real-time speaker diarization, long-context disambiguation, and coverage of a broader range of application scenarios. QwenCloud describes it as building on Qwen-Omni's infrastructure and multimodal data.
It is served only as an API, under the model id `qwen3.8-livetranslate-flash-realtime`, through the WebSocket Realtime API of Alibaba Cloud Model Studio and on QwenCloud. The context window is 53,248 tokens, with up to 49,152 input and 4,096 output tokens, and a limit of 10 requests and 100,000 tokens per minute. Function calling, structured outputs, web search, fine-tuning, batch inference and context caching are not supported.
In the Singapore (international) region it lists $7.50 per 1M audio-input tokens, $0.55 per 1M image-input tokens, $20.00 per 1M text-output tokens and $30.00 per 1M audio-output tokens. In China (Beijing) the same items are $5.653, $0.466, $14.133 and $22.613.
| Released | 2026-09-19 |
|---|---|
| License | Proprietary (hosted API) |
| Weights | API only |
| Parameters | Not disclosed |
| Context | 53,248 tokens |
| Max output | 4,096 tokens |
| Modalities | Audio, Image, Text |
Pricing
| Input | $7.50 / 1M audio tokens |
|---|---|
| Output | $30.00 / 1M audio tokens |
Singapore (international) rates: audio input $7.50, audio output $30.00, image input $0.55, text output $20.00 per 1M tokens. China (Beijing): $5.653 / $22.613 audio, $0.466 image in, $14.133 text out.
Strengths
- Understands 60 languages and produces speech in 29 of them
- Average lagging (LAAL) of 2.3 s across 60 languages, down from 2.8 s, per the Qwen launch post
- Real-time speaker diarization and synchronized display of source text and translation
- Accepts images alongside audio, and returns both text and audio in one stream
- 53,248-token context for long sessions, with long-context disambiguation of names and terms
Best for
- Reach for it for live interpretation of meetings, talks and broadcasts where listeners need translated speech with little delay
- Reach for it for bilingual subtitles that show the source text and its translation together
- Reach for it for multi-speaker conversations where each translated line should be attributed to its speaker
- Look elsewhere when you need self-hosted weights, function calling, structured outputs or batch translation of files, which this API-only model does not offer
How to access
| Provider | Model ID |
|---|---|
| Alibaba Cloud Model Studio (WebSocket Realtime API) ↗ | qwen3.8-livetranslate-flash-realtime |
| QwenCloud ↗ | qwen3.8-livetranslate-flash-realtime |
FAQ
What is Qwen3.8-LiveTranslate?
It is Alibaba's real-time simultaneous interpretation model, announced by the Qwen team on 19 September 2026. It takes live speech (and optionally images) and returns translated text and speech as the speaker talks, understanding 60 languages and speaking 29.
How fast is it?
The Qwen launch post reports an average lagging (LAAL) of 2.3 seconds across 60 languages, down from 2.8 seconds.
What does it cost?
In Alibaba Cloud's Singapore region it lists $7.50 per 1M audio-input tokens, $30.00 per 1M audio-output tokens, $0.55 per 1M image-input tokens and $20.00 per 1M text-output tokens. China (Beijing) pricing is lower: $5.653, $22.613, $0.466 and $14.133.
Are the weights open?
No. The model is offered only as a hosted API, as qwen3.8-livetranslate-flash-realtime on Alibaba Cloud Model Studio's WebSocket Realtime API and on QwenCloud.