Alibaba's Qwen Team Unveils Qwen3.8-LiveTranslate: A Real-Time Translation Model Reducing Average Latency to 2.3 Seconds in 60 Languages

If you build live translation tools, host multilingual meetings, or stream content across languages, Qwen's newest release deserves your attention. The Alibaba Qwen team has launched Qwen3.8-LiveTranslate, a real-time simultaneous interpretation model that listens to live speech — with optional video frames — and produces translated text and voice while the speaker is still talking. The headline improvement is a drop in average lag from 2.8 seconds to 2.3 seconds, alongside new features for multi-speaker sessions, bilingual captions, and long-conversation consistency. Here's what changed, what it can do, and what it costs to run.

What Changed Under the Hood

Simultaneous interpretation is fundamentally a balancing act. The longer the model waits, the more context it has to work with — but every second of waiting is a second of delay for the listener. Qwen3.8-LiveTranslate restructures this tradeoff with a new Interleave architecture that the company says improves faithfulness, fluency, and conciseness all at once.

The latency benchmark to know here is LAAL, or Length-Adaptive Average Lagging. This metric measures how far behind the translation typically trails the original speech, and it's designed so systems can't game the score by producing excessive output. Cutting average lag from 2.8 seconds to 2.3 seconds works out to roughly an 18% reduction.

According to the QwenCloud team, this model is the real-time counterpart to Qwen3.8-LiveTranslate-Flash. It builds on the Qwen-Omni stack, trained with large-scale multimodal data, cross-language and cross-modal alignment, plus visual enhancement. The Flash variant also handles offline audio and video translation if you don't need live streaming.

Three New Capabilities

Real-time speaker diarization. In multi-party conversations, the model can tell speakers apart and keep each person's voice consistent through more stable voice cloning. Developers get access to cloning modes through the API, including an "always" mode that re-clones the voice before every response — useful when several people are speaking in the same session.

Synchronized bilingual display. Viewers see the source text and its translation side by side. On the technical side, the source transcription streams as its own separate event stream alongside the translation stream.

Long-context disambiguation. The model draws on conversation history to resolve names and terminology. If someone introduces a person or product name early in a meeting, the translation stays consistent with it hours later.

Languages, Inputs, and Vision

The model understands 60 languages and can speak 29 of them, returning both audio and text. The remaining 31 languages return text only. Voice output includes Chinese, English, Arabic, German, French, Spanish, Japanese, Korean, Hindi, and more.

Input comes as audio plus optional images. Visual cues — lip movements, gestures, on-screen text — help the model in noisy rooms or when words are ambiguous. The documentation advises sending no more than 2 images per second.

Teams can also configure hotwords, mapping source terms to fixed target translations. The docs suggest keeping this list at 1,000 entries or fewer.

API, Pricing, and Limits

Developers connect through the WebSocket Realtime API using the model ID `qwen3.8-livetranslate-flash-realtime`, available on Alibaba Cloud Model Studio and QwenCloud. The default turn detection type is `speaker_detection`, and clients stream audio continuously while receiving server-generated responses.

Default audio formats are 16 kHz PCM for input and 24 kHz PCM for output, with Tina as the default voice. You can set `session.output_modalities` to text-only or text-plus-audio. One important detail: always send `session.finish` before closing the connection, or the final translated segment gets lost.

Singapore list pricing per 1M tokens:

• Audio input: $7.50
• Image input: $0.55
• Text output: $20
• Audio output: $30

Beijing-region pricing is lower, at $5.653, $0.466, $14.133, and $22.613 respectively. In token terms, audio input consumes 7 tokens per second and audio output consumes 12.5 tokens per second. An hour of two-way speech works out to roughly $1.54 in Singapore, before accounting for text and image tokens.

The context window is 53,248 tokens — 49,152 for input and 4,096 for output. Default rate limits are 10 requests and 100,000 tokens per minute. Note that Model Studio lists function calling, structured outputs, batch inference, and fine-tuning as unsupported.

Key Takeaways

• Average lag (LAAL) drops from 2.8s to 2.3s with the new Interleave architecture
• New features include speaker diarization, bilingual display, and long-context disambiguation
• 60 languages understood, 29 spoken aloud
• Access is API-only via Alibaba Cloud Model Studio and QwenCloud