Tencent Hunyuan's real-time simultaneous interpretation model, supporting bidirectional Chinese-English translation: it continuously receives audio and outputs translated text as it listens. Once a TTS service is configured, it can also output translated speech that preserves the speaker's voice. This is the FP8 quantized version; the series also includes BF16, GGUF, and an Interleave version that outputs both the original and translated text.
Rising score 132 · Likes 135 · Downloads 93
Reasons
- Officially, in the RealSI long-audio evaluation, StreamLAAL (streaming translation latency) is approximately 3 seconds, the lowest among compared systems. The GEMBA quality score is higher than that of the compared open-source models and 4 out of 5 commercial APIs. The official documentation also notes that StreamLAAL is not equivalent to time-to-first-audio.
- Officially, on four short-audio benchmarks: FLEURS 90.3, CoVoST2 75.5, ACL6060 81.2, and RealSI-Sentence 84.7 (the highest in the table). FLEURS is lower than Gemini3.5-LiveTranslate's 92.4, and CoVoST2 is lower than Qwen3.8-LiveTranslate's 82.1.
- Officially, the model was compared against 6 simultaneous interpreters holding Level 1 and 2 CATTI certificates using 18 segments of new Chinese-English audio released after 9 2026 (each segment approximately 6 minutes long). The official claim is that quality is comparable and latency is often lower, but it is noted that this conclusion applies only to this specific sample set. Released on 26 10, it was not included in previous editions; the rising score of 132 only reflects attention level.
Usage
- One-click deployment: git clone https://github.com/Tencent-Hunyuan/Hy-SimulMT.git，进入 SimulInfer/ and run bash deploy.sh --install-server --model tencent/Hy-SimulMT-30B-A3B-Interleave (the --model argument can be replaced with IDs of other versions or local paths). Once 'Ready' appears, open http://127.0.0.1:8080/，选翻译方向，上传音频或开麦克风。.
- Command line: curl -N http://127.0.0.1:8080/translate/stream -F direction=en-zh -F [email protected]. To output translated speech and voice cloning, first run cp deploy.example.toml deploy.toml, configure the TTS connection within it, and then run bash deploy.sh --config deploy.toml.
- Environment: Linux, Python 3.12+, NVIDIA GPU. Existing environments require vLLM >=0.31 (supporting HYV3ForCausalLM) and simulst installed. Alternatively, SGLang-Omni can be used (supports only the BF16 version, outputs text only, with a maximum audio length of 30 seconds per request). For the GGUF version, wait for llama.cpp to merge PR #31198.
Risks
- Supports only Chinese-English translation; other languages listed in the tags are not explained in the model card. Translated speech and voice cloning depend on separately configured TTS services.
- The model card does not specify the VRAM requirements for the FP8 version. Officially, the BF16 version weights are approximately 57 GiB (excluding KV cache), which is sufficient for a single H200 GPU.
- Scores and interpreter comparisons are official self-tests, and the interpreter comparison sample size is small. No similarity data is provided for voice cloning. The two demo videos are offline replays, and the official documentation states they cannot be used for latency measurement.
Editor's note: The GitHub repository includes the SimulTrain training tool, which allows continued training from published checkpoints. If only text translation is needed, the SGLang-Omni path is lighter but does not cover the FP8 version.