Alibaba has launched Qwen3.8-LiveTranslate, a real-time interpretation model designed to reduce latency in simultaneous translation. The system utilizes an Interleave architecture to balance context gathering and response speed, achieving an average lag of 2.3 seconds, which represents an 18% improvement over the previous 2.8-second metric.
The model supports 60 languages for input, with 29 of those available for audio and text output. It processes audio and optional image inputs, with a recommended limit of 2 images per second. Developers can integrate the model via the WebSocket Realtime API on Alibaba Cloud Model Studio and QwenCloud using the identifier qwen3.8-livetranslate-flash-realtime.
Technical specifications include a context window of 53,248 tokens, split between 49,152 for input and 4,096 for output. The system operates with default rate limits of 10 requests and 100,000 tokens per minute. Audio input is processed at 16 kHz, while output is generated at 24 kHz using the Tina voice profile.
Pricing varies by region, with costs in Singapore estimated at $1.54 per hour for speech-to-speech processing. Users can configure up to 1,000 hotwords for specific translations. The platform currently does not support features such as fine-tuning, batch inference, structured outputs, or function calling.
