Tencent open-sourced Hy-SimulMT-30B-A3B-Interleave for low-latency Chinese-English simultaneous translation
Tencent released Hy-SimulMT-30B-A3B-Interleave, a real-time simultaneous translation model, on Hugging Face under Apache-2.0. It supports bidirectional Chinese-English speech-to-speech translation with voice cloning. The model achieved a StreamLAAL of approximately 3 seconds on the RealSI benchmark.
China context
- Original name
- 腾讯混元同声传译模型
- Outside China
- Open weights · huggingface.co
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Builders outside China can integrate Hy-SimulMT via Hugging Face and the provided SimulInfer service, with Apache-2.0 permitting commercial use. The model's low latency and voice cloning enable real-time translation features in applications.
- For investors
- Tencent's release of a competitive open-weights simultaneous translation model may pressure commercial translation API providers and accelerate open-source adoption in real-time communication tools.
Translated from Chinese. Quotes and facts link to the original sources.
Tencent Hunyuan released Hy-SimulMT-30B-A3B-Interleave, a low-latency simultaneous translation model for Chinese-English speech-to-speech translation. The model is available on Hugging Face under Apache-2.0 and includes real-time voice cloning. On the RealSI long-form audio benchmark, it achieved a StreamLAAL of approximately 3 seconds, the lowest among compared systems, and its GEMBA score exceeded those of compared open-source models and four of five commercial APIs. The GitHub repository provides SimulInfer inference service and SimulTrain training tools; speech output and voice cloning require a configured TTS service. The model can be tried in the Hy Translate App.
Hy-SimulMT-30B-A3B-Interleave is a 30B-parameter model with 3B active parameters (A3B) using an interleaved architecture, optimized for streaming translation. It processes continuous audio input and generates translated text and speech with low latency. The model's StreamLAAL of approximately 3 seconds indicates minimal lag in streaming translation. The repository includes SimulInfer for inference and SimulTrain for training, suggesting a complete pipeline for simultaneous translation development.
Developers outside China can now access a competitive open-weights simultaneous translation model under Apache-2.0, reducing the cost of building real-time translation applications. Tencent's release directly challenges commercial APIs by offering comparable quality at lower latency, potentially shifting adoption toward open-source solutions.
Hy-SimulMT-30B-A3B-Interleave provides a cost-effective solution for real-time translation in meetings, speeches, and interviews. Its open-source license allows commercial use without licensing fees, and the low latency improves user experience. The model's voice cloning capability adds value for applications requiring speaker identity preservation.
Watch for independent evaluations of Hy-SimulMT on additional benchmarks and real-world scenarios to verify the claimed performance. The release of training tools may lead to community fine-tuned variants for other language pairs. Adoption in live communication tools and integration with TTS services will be key indicators of impact.