On this page

On October 26, 2026, Tencent Hunyuan’s 2026.10.26 | Release note opened Hy-SimulMT: bidirectional Chinese↔English speech-to-speech simultaneous translation with real-time voice cloning. The project ships on GitHub as Tencent-Hunyuan/Hy-SimulMT, with weights on Hugging Face as tencent/Hy-SimulMT-30B-A3B. The Hy-SimulMT license file states Apache-2.0 for Tencent’s inference code, parameters, and weights (third-party components keep their own licenses). Safetensors total about 30.4B parameters (30,394,794,496); the card name is Hy-SimulMT-30B-A3B (FP8 / GGUF / Interleave variants are also mentioned around the release). Same-lab open work on DigIntoAI includes Hunyuan Opens Hy-Nomas: Interactive Video World Model on a 4090.

The pitch is live meetings, speeches, and interviews: keep translating while audio is still arriving, instead of waiting for a full utterance. That is useful for teams that want a self-hosted ZH↔EN path—once the gap between what is open and what still needs wiring is clear. In practice, “open” here means Apache-2.0 code plus downloadable ~30B weights and documented SimulInfer / SimulTrain entry points; it does not mean a finished booth appliance with mic, speaker, and cloned voice already glued together. TTS and audio I/O still have to be wired on the deployer’s side.

~3 seconds is StreamLAAL—not first speech

On the RealSI long-form audio evaluation, Hunyuan reports a StreamLAAL of approximately 3 seconds, the lowest among the systems shown in their comparison figure. README text is explicit in both English and Chinese: StreamLAAL is not the same as the time to first speech output. StreamLAAL measures streaming translation lag along the horizontal axis of their GEMBA-vs-StreamLAAL plot; it is not a promise that the first audible translated syllable arrives in three seconds. If a vendor slide or second-hand post collapses those into “3-second first speech,” that is a misread of the README.

GEMBA scores, the short-form tables (FLEURS, CoVoST2, ACL6060, RealSI-Sentence), and the matched comparison with CATTI-certified interpreters are vendor-reported. The interpreter study uses 18 newly collected Chinese and English recordings (about six minutes each) across ten domains, with six interpreters (three CATTI Level 1, three Level 2). Hunyuan’s own README frames those findings as applying only to the samples, participating interpreters, and evaluation settings used—and says they do not constitute a comprehensive assessment of speech quality, voice similarity, or other dimensions. This piece keeps that same limit and does not frame the work as replacing human interpreters.

Download ≠ ready to interpret out loud

The repository provides SimulInfer (inference service) and SimulTrain (training tools). The README states plainly that speech output and voice cloning require a configured TTS service. Open weights give the translation / streaming stack; they do not by themselves speak. The product path in the Hy Translate App is also available, but that is not the same as cloning a self-hosted pipeline from GitHub and Hugging Face.

Two deployment edges matter for local teams:

  1. llama.cpp is not merged yet. Hunyuan says it filed PR #31198. As of October 27, 2026 (Beijing time), that PR was still open and unmerged, with current work oriented toward llama-mtmd-cli (llama-server described as later). Official llama.cpp does not yet ship Hy-SimulMT support.
  2. Hugging Face has no hosted inference provider for this card. The API inferenceProviderMapping was empty; the model page wording is that the model is not deployed by any Inference Provider. Pulling weights and serving them on a self-hosted GPU stack is the open path—not a one-click HF provider endpoint.

Open-source speech peers on DigIntoAI for contrast: Block Releases Dessa, a Full-Duplex Speech Model Trained Only on Speech.

Language scope and quality claims

The product release sentence is bidirectional Chinese–English speech-to-speech. Hugging Face language tags may list additional codes (including Japanese). Those tags are not treated here as a claim that Japanese or other conference languages are already supported for simultaneous interpreting on this release. For JA↔EN meetings or a multilingual booth, ZH↔EN remains the confirmed product scope until Hunyuan says otherwise in a release note.

Vendor-reported GEMBA and short-form numbers can help a team decide whether to spend engineer time on a trial—they are not a substitute for that team’s own meeting audio, microphone chain, or domain glossary. Open source also does not erase GPU, TTS, and ops cost: Apache-2.0 removes a license tax, not a deployment bill. A workable pilot checklist usually covers: confirm ZH↔EN is the only required pair; budget a TTS path for audible output; treat llama.cpp support as still a PR, not a release; and measure first-speech delay on the team’s own audio, not StreamLAAL alone.

For choosing among speech models for recordings (transcription, edge, multi-speaker meetings) rather than live ZH↔EN interpreting, DigIntoAI’s Granite guide is a different job: Which IBM Granite Speech Model Fits Your Recording?.

End-to-end latency, voice quality, cloning similarity, and self-host cost on Hy-SimulMT have not been tested on this site; feedback can go through the site-owner contact link in the footer.