Tencent releases Hy-SimulMT-30B-A3B for real-time translation

Tencent published Hy-SimulMT-30B-A3B, an open-weight model for real-time simultaneous translation, on Hugging Face under the Apache 2.0 license. The model card describes it as a low-latency, high-quality model built for simultaneous translation, meaning it can translate speech or text as it comes in rather than waiting for a full sentence or passage to finish.

The model has about 30 billion total parameters and supports a wide set of languages, including Chinese, English, French, Portuguese, Spanish, Japanese, Turkish, Russian, Arabic, Korean, Thai, Italian, German, Vietnamese, and dozens more listed on its model card, covering nearly 40 languages in total. It uses the Transformers library and is set up for text generation with a translation pipeline.

Tencent's own model card is the only detail available on the model's architecture and intended use at this point; no separate benchmark results were captured in this source.

Why it matters

An open, low-latency translation model could let developers build real-time multilingual tools without relying on closed translation APIs.

Source: Tencent on Hugging Face