Hy-SimulMT 30B-A3B

Not yet independently verified. Not yet independently verified. The release date and the latency and quality comparisons come from Tencent's own repository; we found no independent test, and the licence is read from the model card metadata. We will update this when it can be confirmed, and remove this note.

Hy-SimulMT is Tencent Hunyuan's model for simultaneous translation between Chinese and English, in both directions. Its repository dates the release to 26 October 2026. Rather than waiting for a sentence to finish, it keeps listening to the incoming audio and translates as the speaker goes, which is what an interpreter does in a meeting, a speech or an interview.

Translating before the sentence ends

The model outputs translated text as the audio arrives. Spoken output in the speaker's own voice, which Tencent calls real-time voice cloning, needs a text-to-speech service configured alongside it; the open release is the translation model. Tencent's GitHub repository supplies an inference service, SimulInfer, and training tools, SimulTrain, for fine-tuning. The weights are published in three forms: the base 30B-A3B model, an FP8 quantised version and a GGUF build for local runtimes. The model card metadata gives the licence as Apache 2.0.

How Tencent measures it

Tencent tested the model on RealSI, an open benchmark of long-form audio, and reports a StreamLAAL latency of about three seconds, the lowest of the systems in its comparison, with a GEMBA translation-quality score above the open-source models it tested and above four of the five commercial translation APIs. In a small assessment against human simultaneous interpreters holding China's CATTI level 1 and level 2 certificates, Tencent says quality was on a par and the model often responded sooner. These are Tencent's own evaluations and are not stored as scores.

One language pair, both directions

The model covers Chinese and English only, so it suits businesses with regular meetings or events across that pair rather than general multilingual captioning. Tencent also offers it to consumers through its Hy Translate app.

Key facts

Maker: Tencent (Hunyuan team) Released: 26 October 2026 Languages: Chinese and English, both directions Size: about 30B parameters, about 3B active Formats: BF16, FP8 and GGUF Licence: Apache 2.0 (model card metadata) Board family: Tencent Hy