model releaseTencent

Tencent Hunyuan open-sources Hy-SimulMT, a 30B-A3B speech-to-speech interpreter with about 3 seconds of lag

TL;DR

Tencent Hunyuan has open-sourced Hy-SimulMT, a 30B-A3B mixture-of-experts model that translates Chinese and English speech in real time while cloning the speaker's voice. Tencent reports about 3 seconds of StreamLAAL lag on the long-form RealSI benchmark, the lowest among the systems it compared. The weights are listed under Apache 2.0.

3 min read
0

Tencent Hunyuan has open-sourced Hy-SimulMT, a simultaneous interpretation model that reports about 3 seconds of translation lag on the long-form RealSI benchmark. It converts Chinese speech to English speech, and English to Chinese, while the speaker is still talking, according to Pandaily, which cites IT Home and the project's GitHub and Hugging Face pages.

The model clones the speaker's voice in real time, so listeners hear the translation in a voice close to the original. It is available first in the Hy Translate app, which is free to use.

Latency and quality

Tencent measures lag with StreamLAAL, a standard metric for streaming translation delay. It reports about 3 seconds on RealSI, the lowest among the systems it compared. The project notes that StreamLAAL is not the same as the time until the first translated speech is heard. IT Home puts the typical lag of professional human interpreters at 4 to 5 seconds.

On quality, Tencent says Hy-SimulMT's GEMBA score on long-form speech beat the open models in the comparison and four of the five commercial APIs tested. On short-form sets it scored:

  • FLEURS: 90.3
  • CoVoST2: 75.5
  • ACL6060: 81.2
  • RealSI-Sentence: 84.7, the highest of every listed system

Comparison with human interpreters

The team also compared the model with human interpreters on 18 recordings of about six minutes each, recorded from September 2026 onward across 10 domains. Six interpreters took part: three holding China's top CATTI Level 1 interpreting certificate and three holding Level 2. The best result from each level was used. Tencent calls this a small-scale assessment.

Wang Weiwei, vice dean of the School of Interpreting and Translation Studies at Guangdong University of Foreign Studies, said in remarks published with the release that the system speaks first and corrects itself afterward, much like a human interpreter. He also said AI interpreting is moving from helping people understand toward usable communication in low-risk settings.

Model and tooling

  • Architecture: 30B-A3B mixture-of-experts
  • Builds: FP8 and GGUF, plus an Interleave variant that also outputs the source transcript
  • License: Apache 2.0, as listed on the Hugging Face card
  • SimulInfer: a serving stack built on vLLM, with a web interface, a WebSocket API and optional text-to-speech
  • SimulTrain: a tool for further training on domain data
  • llama.cpp: support has not yet been merged

Context window and API pricing are not yet disclosed. The Hy Translate app is free.

The release follows Hunyuan's earlier open translation models and moves the line from text into live speech.

What this means

The headline figure is lag, but it needs careful reading. StreamLAAL measures streaming translation delay, not the wait before the first translated speech is heard, so the 3-second figure does not describe that initial pause.

The Apache 2.0 license, FP8 and GGUF builds, and the SimulInfer and SimulTrain tools mean the weights, a serving stack and a fine-tuning path are all included. Teams can host the model themselves and adapt it to domain data. The missing llama.cpp support is a practical limit for anyone planning to run the GGUF builds through that runtime.

The human comparison covers 18 recordings and six interpreters, and Tencent itself describes it as small-scale. Wang Weiwei's framing points to low-risk settings as the realistic target for this kind of system.

Related Articles

product update

Tencent adds three video models to WorkRally: Rally-Video, Rally-Video-edit-2.0 and UpScale-rally with 4K output

Tencent has upgraded its WorkRally AI content-creation platform with three video models: Rally-Video for generation, Rally-Video-edit-2.0 for selective editing, and UpScale-rally for 2K and 4K super-resolution. Tencent claims UpScale-rally cuts processing time by 80% versus the previous version. Pricing was not disclosed.

model release

Taobao's MNN team open-sources MNN-Pocket-0.3B, a 300M-parameter on-device model with a 173MB INT4 package

Alibaba's MNN team at Taobao has open-sourced MNN-Pocket-0.3B, a roughly 300-million-parameter model built to run inside mobile apps. The INT4 deployment package is 173MB, and the team reports about 350 tokens per second decoding on an iPhone 18 Pro.

changelog

Tencent expands Hy4 inference cluster tenfold after WorkBuddy usage surge

Tencent says it has expanded the inference cluster serving Hy4 tenfold after many WorkBuddy users hit their quotas on the model's first day. Hy4 scores 35 on the Artificial Analysis Intelligence Index, second among Chinese models, and ranks 13th on Agent Arena.

model release

Nous Research releases Arete-27B, a Qwen3.8-27B fine-tune that cuts propaganda rate from 22% to 0.9%

Nous Research fine-tuned Alibaba's Qwen3.8-27B on about 53 million tokens to remove Chinese state propaganda from its answers. According to Nous, the resulting Arete-27B model drops the propaganda rate from 22 percent to 0.9 percent while keeping general capabilities intact. Weights, test set and training data are public under Apache 2.0.

Comments

Loading...