On October 26, 2026 (China time), Hunyuan, the team that develops AI at China's Tencent, released Hy-SimulMT, a simultaneous interpretation AI model. It starts translating while the speaker is still talking and outputs the translation as text and speech. It works between Chinese and English in both directions and also supports "voice reproduction (voice cloning)," which produces the translated speech in a voice resembling the speaker's. The model weights are available on Hugging Face and elsewhere under the Apache-2.0 license, which is friendly to commercial use.
What is Hy-SimulMT?
Hy-SimulMT is designed for live exchanges such as meetings, lectures and interviews. It receives audio and keeps producing translations as it goes.
The challenge of simultaneous interpretation is that waiting until the speaker finishes makes the meaning easier to grasp, but keeps listeners waiting. Tencent says it developed the model with an emphasis on balancing translation quality and speed. An interpreting expert from Guangdong University of Foreign Studies who took part in the evaluation said the way it starts translating early to keep the delay down and then rephrases itself resembles the self-corrections of a human interpreter.
Performance
Tencent compared the model with other companies' open models and commercial APIs on RealSI, a test that evaluates simultaneous interpretation on long audio.
- Its delay metric (StreamLAAL) was about 3 seconds, the shortest among the systems compared. However, this metric is different from the time until the first audio is output
- Its translation quality score (GEMBA) beat every open model compared and four of the five commercial APIs
- On four tests using short audio, it scored 84.7 on RealSI-Sentence, which translates sentence by sentence, the highest score in the table
Tencent also compared it with six interpreters holding level 1 or level 2 of CATTI, China's translation qualification, using 18 Chinese and English audio clips (about 6 minutes each) published online since September 2026. The clips were lectures and meetings across 10 fields, including software, medicine, law and education. Tencent says the model translated with quality comparable to that of professional interpreters, and in many cases faster. However, it notes that this comparison was small and the results are limited to these clips and conditions. All of these are Tencent's own evaluations.
Translating in the speaker's own voice
Hy-SimulMT can produce speech of the translated words while keeping the characteristics of the original speaker's voice. Tencent says this makes it easier to tell who is speaking even when several people are talking. The official page includes demos such as translating English audio of three people talking into Chinese.
How to use it and hardware requirements
The released model is named Hy-SimulMT-30B-A3B. The variants are as follows.
- A model that outputs only the translation, and an "Interleave" version that outputs the transcript of the original speech together with the translation
- For each, a lighter FP8 version and a GGUF-format version
On GitHub, Tencent has also released SimulInfer for running the model and SimulTrain for training it. SimulInfer has a web interface where you can try it by uploading an audio file or speaking into a microphone. The standard setup requires a Linux machine with an NVIDIA GPU, and the weights take up about 57 GiB. Support in llama.cpp for running the GGUF versions has not yet been merged into the official project. To output the translation as speech or reproduce voices, you need to set up a separate text-to-speech (TTS) system.
For general users, Tencent says it can also be tried in its translation app, Hy Translate App (Android and iOS). We could not confirm on the official page whether it can be used from Japan.
CoAI's take
Simultaneous interpretation for online meetings and lectures is one of the most anticipated uses of AI. What stands out about Hy-SimulMT is that it delivers a short delay and voice reproduction in a form anyone can run on their own hardware. In the short-audio tests, it was also compared with simultaneous interpretation features from Qwen, Gemini, OpenAI and others.
However, so far it officially claims support only for Chinese and English, and use with Japanese is not envisaged. It is worth trying for dealing with partners in Chinese-speaking regions or listening to English lectures, but its performance is backed mainly by Tencent's own evaluations. Tencent itself says professional interpreters remain important in high-stakes situations or where expert review is needed. Because the feature can also reproduce other people's voices, it should be handled with care, such as by getting the other person's consent before use.