SSyncDev

Hy-SimulMT-30B-A3B

SyncDev model directory · Profile updated Nov 17, 2026

Compare with another model

Hy-SimulMT-30B-A3B is Tencent's open-weight Apache-2.0 MoE for simultaneous Chinese-English speech translation, with about 3 seconds of reported latency.

Key specifications

Specification score
9.9
License
Apache-2.0

The specification score summarizes recorded features and limits. It is not a benchmark result.

Inputs and outputs

Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.

Accepts

  • Audio input

Produces

  • Text output

Capabilities

AudioOpen weights

What is Hy-SimulMT-30B-A3B?

A human interpreter starts talking before the speaker has finished. Most machine translation waits for the full sentence. Tencent's Hy-SimulMT closes that gap: a mixture-of-experts model that listens to a continuous audio stream and emits the translation as it goes, between Chinese and English in both directions.

The weights went up on 26 October 2026 under Apache-2.0. There is no hosted API from Tencent for this one; you run it yourself.

Correcting a common description

Some early write-ups describe it as text-to-text. The official model card says otherwise. Input is continuous audio. Output is translated text, and speech only if you attach a TTS service. The card also describes optional real-time voice cloning for spoken output. Plan your pipeline around audio in.

The card's metadata lists 36 languages, but the documented capability is Chinese and English only. Do not expect other pairs.

Size and builds

The name encodes the shape: about 30B parameters in total, roughly 3B active per token. That active count is inferred from the "A3B" naming; the card does not state it directly.

Build Precisions What it outputs
Translation-only BF16, FP8, GGUF Translation
Interleave (tencent/Hy-SimulMT-30B-A3B-Interleave) BF16, FP8, GGUF Source transcript and translation together

Choose Interleave if you need live captions in both languages side by side.

Running it

Three routes are documented:

  • SimulInfer, via a deploy script. Needs Python 3.12+, an NVIDIA GPU and vLLM 0.31 or later with support for the model's architecture. It serves a web/API front end on port 8080.
  • llama.cpp, through a pull request that had not been merged when we checked, using the multimodal CLI. Server support is planned.
  • SGLang-Omni, which covers only the BF16 translation-only checkpoint and returns text. A single H200 is reported to be enough, with up to 6.5x speed-up over Transformers at batch size one.

A training toolkit, SimulTrain, lets you continue training from the released checkpoint.

Reported quality and latency

Tencent reports around 3 seconds of streaming lag (StreamLAAL) on a long-form benchmark, the lowest among the systems it compared. On short-form sets it lists FLEURS 90.3, CoVoST2 75.5, ACL6060 81.2 and RealSI-Sentence 84.7. Competing systems scored higher on FLEURS and CoVoST2. A small human study on eighteen six-minute recordings found quality on par with expert interpreters; Tencent itself flags how small that study is.

Trying it on a real meeting

Record a bilingual meeting with crosstalk and domain jargon. Stream it through the model and log the delay between each source phrase and its translation. Then have a bilingual colleague mark errors in names, numbers and negations, the three things simultaneous systems most often get wrong.

Other options are under translation models and speech recognition. Tencent's wider catalogue, including Hy3, is on its provider page.

How to evaluate Hy-SimulMT-30B-A3B

Compare the model on the work you need to complete.

Tasks and inputs

Hy-SimulMT-30B-A3B is listed for Audio, Automatic speech recognition, Translation. It accepts audio and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.

Capacity and cost

Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.

Choosing a deployment

The weights are downloadable under Apache-2.0. Check runtime support and size the complete model and working context for your expected concurrency.

Frequently asked questions

Continuous speech between Chinese and English in both directions, producing translated text as the speaker talks; spoken output needs an added TTS service.

These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.

Alibaba

Qwen3.8 LiveTranslate Flash Realtime

Qwen3.8 LiveTranslate Flash Realtime translates live audio and visual context, recognizing 60 source languages with voice output in 29.

— context

Tencent

Hy-MT2-7B

Hy-MT2-7B from Tencent: Hy-MT2-7B translates text between supported languages.

8.19K context

Alibaba

Qwen MT Plus

Qwen MT Plus from Alibaba: Qwen MT Plus translates text between supported languages.

16.38K context