Tencent opens simultaneous speech translation
On 26 October 2026 Tencent's Hunyuan team released Hy-SimulMT, a 30B-A3B mixture-of-experts model that translates a continuous audio stream between Chinese and English while the speaker is still talking, with its weights and the SimulInfer inference service and SimulTrain training tools published under Apache 2.0.
Why it matters
Simultaneous interpretation, so far a human skill and a feature of closed services, is released as open weights for one language pair. The released model takes audio in; speech output and the voice cloning that the announcement advertises need a separate text-to-speech service that the user must configure, so the open part is the translation and not the whole voice-preserving interpreter.
What the pages say. The Hugging Face card and the GitHub README both carry the line "2026.10.26 | Release: Hy-SimulMT is released with bidirectional real-time simultaneous translation between Chinese and English and real-time voice cloning." The repository provides the SimulInfer inference service and SimulTrain training tools, and states "Speech output and voice cloning require a configured TTS service". The weights are 17 safetensors shards plus an audio tower, 30.4 billion parameters in all. Licence: LICENSE.txt says Hy-SimulMT is licensed under Apache-2.0, except for third-party components listed in it. The model can also be tried in Tencent's Hy Translate app. Dates. The Hugging Face repository was created on 22 October 2026 and has commits from then; the release line is dated 26 October and is the date of the record. Whether the repository was public earlier is not shown. What the record does not claim. The card compares the model with open-source systems and commercial APIs on the RealSI long-form audio benchmark (a lag of about 3 seconds, which it calls the lowest of the systems shown), in a table of scores on four short-form sets, and with six human interpreters; these are Tencent's own evaluations and none is repeated as a finding. The README speaks of preserving the speaker's vocal characteristics; that depends on the separate speech service and was not tested.