Alibaba’s Qwen team released Qwen3.8-LiveTranslate on September 19, 2026, introducing a next-generation model for real-time simultaneous interpretation. It listens to live speech, uses optional video frames, and returns translated text and speech while the speaker is still talking.
The main change is a new Interleave architecture, which rebuilds the interpretation loop around the tension at the heart of live translation: waiting for more context can improve accuracy, but speaking sooner keeps the conversation moving. Qwen says the new model improves faithfulness, fluency, and conciseness while cutting average lag from 2.8 seconds to 2.3 seconds.
A shorter wait without losing the conversation
Qwen measures this delay with average lagging, or LAAL. The measure tracks how far a translation trails the original speech on average, making it useful for a system that must listen and respond during an ongoing conversation. The drop from 2.8 seconds to 2.3 seconds represents roughly an 18% reduction in average lag.
That change matters because simultaneous interpretation cannot wait for every sentence to end before producing an answer. The model must decide when it has enough information to translate, then keep updating its understanding as more speech arrives. The Interleave architecture is designed around that live exchange between incoming speech and outgoing translation.
Qwen3.8-LiveTranslate also adds real-time speaker diarization, synchronized bilingual display, and long-context disambiguation. Speaker diarization helps separate speakers during a live session, while the bilingual display keeps the original and translated content aligned. Long-context disambiguation helps the model use more of the surrounding conversation when a word or phrase has more than one possible meaning.
Audio, video, and 60-language support
The model builds on the Qwen-Omni stack, large-scale multimodal data, cross-language and cross-modal alignment, and visual enhancement. Its inputs are audio and optional images, while its outputs are text and audio. The Flash model also supports offline audio and video translation.
Qwen3.8-LiveTranslate understands 60 languages and can speak 29 of them, returning audio along with text. Visual information can add useful clues when sound alone is not enough. Lip movements, gestures, and on-screen text can help in noisy rooms or when spoken words are ambiguous.
Teams can also set hotwords for terms that need special attention, with a recommended maximum of 1,000 hotwords. This gives developers a way to guide recognition for important vocabulary within a live interpretation workflow.
API access, limits, and pricing
Qwen3.8-LiveTranslate is available only through an API. It is live on Alibaba Cloud Model Studio and QwenCloud as qwen3.8-livetranslate-flash-realtime, with developers connecting through the WebSocket Realtime API.
The default rate limits are 10 requests and 100,000 tokens per minute. The model offers a context window of 53,248 tokens, including 49,152 input tokens and 4,096 output tokens. Model Studio lists function calling, structured outputs, batch inference, and fine-tuning as unsupported.
Pricing depends on the type of data moving through the system and the service location. In Singapore, the price per 1 million tokens is $7.50 for audio input, $0.55 for image input, $20 for text output, and $30 for audio output. In Beijing, the prices are lower at $5.653, $0.466, $14.133, and $22.613 USD respectively.
For a simpler reference point, one hour of speech in and speech out costs about $1.54 in Singapore before text and image tokens. That estimate does not cover any added text or visual processing, so the final cost depends on how a project uses the model.
The release positions Qwen3.8-LiveTranslate as a hosted tool for applications that need translation during live speech rather than after a recording ends. Its faster average response, speaker separation, synchronized bilingual output, visual cues, and long-context handling all focus on one goal: keeping interpretation close to the pace of the conversation.
Based on
