Tencent Built a 3-Second Interpreter and Shipped It Without a Press Release

On 22 October 2026, six repositories in a new model family appeared under the tencent organisation on Hugging Face. The one this article is about is Hy-SimulMT-30B-A3B-Interleave, the variant that emits both the recognised source text and the translation; its sibling Hy-SimulMT-30B-A3B emits the translation alone. Both are Apache-2.0, both are 30.4 billion parameters, and both are built for a job none of Tencent's previous translation releases could do: listening to Chinese or English speech as it is spoken and producing the other language while the sentence is still in progress. Four days later, the main repository showed zero downloads and two likes. There was no Hunyuan blog post, no press release, no arXiv paper, and no line in any of the Chinese-language coverage of Tencent's translation business that had been running all month. The only announcement is a single dated bullet in the model card's News section, added on 26 October, which reads: released.

That gap between what the weights do and how loudly they were published is the story. A simultaneous interpreter with roughly three seconds of lag, downloadable under a permissive licence, is not a routine release — it is the piece that turns a translation app from something you take turns with into something you can wear in a meeting. The likeliest explanation for the silence is on the calendar: Tencent became a strategic partner of the 33rd APEC Economic Leaders' Meeting in Shenzhen on 23 October, the day after the weights went up, and its Hy Translate app was named the official AI translation tool for the volunteers and staff working that event, which runs 18–19 November. That is an inference from public dates, not something Tencent has said. What follows separates what the repository proves from what it merely suggests.

What is actually in the repository

The architecture is a speech-in, text-out causal language model tagged hy_v3 and instantiated as HYV3SpeechForCausalLM. The configuration spells out a sparse mixture-of-experts stack: 48 layers, hidden size 2,048, 128 experts with an expert hidden dimension of 768, eight routed experts plus one shared expert per token, and a single dense layer at the front. The safetensors index counts 30,394,788,480 parameters in BF16, with a further 6,016 stored in F32. Tencent never expands the "A3B" in the name, but the arithmetic is consistent with it: eight routed experts plus one shared, at that expert width, works out to roughly 3 billion active parameters per token. That is our calculation from the published config, not a vendor figure.

The audio front end is in the same config file rather than hidden behind a pipeline: 128 mel bins, a 24-layer encoder at width 1,024 projecting to 2,048, 16 kHz sampling, and a maximum of 480,000 samples — thirty seconds — per input window. Generation defaults to temperature 0.0 with max_new_tokens of 64, which is what you would expect from a model designed to emit a short stretch of translated text and then wait for more audio. Maximum position embeddings are listed as 262,144; the card publishes no usable context length, and for streaming audio that number is far less meaningful than the thirty-second window.

Six checkpoints shipped in one go: base and Interleave, each with an FP8 build and a GGUF build. The licence file is Apache-2.0 with the usual carve-out that bundled third-party components keep their own terms. The model card names twelve contributors, links a GitHub repository containing both an inference service called SimulInfer and a training toolkit called SimulTrain, and cites a Tencent product page rather than a paper.

The latency number, and the benchmark row that undercuts the headline

The claim that matters for a simultaneous model is not translation quality in isolation — it is quality at a given lag. Tencent evaluates this on RealSI long-form audio and reports a StreamLAAL of approximately three seconds, described as the lowest among the systems compared, with a GEMBA score above the open-source models in the comparison and four of the five commercial APIs. StreamLAAL measures the average lag between source and target tokens in a streaming setting; it is not time-to-first-audio, and the card says so explicitly. Nor does the card name the fifth commercial API that beat it on GEMBA.

The short-form table is more informative because it is less flattering in one column. On four public benchmarks, Tencent reports Hy-SimulMT at 90.3 on FLEURS, 75.5 on CoVoST2, 81.2 on ACL6060 and 84.7 on RealSI-Sentence. Against that, Qwen3.8-LiveTranslate posts 88.4, 82.1, 80.5 and 84.4; Gemini3.5-LiveTranslate posts 92.4, 74.4, 82.6 and 82.2; Seed-LiveInterpret2.0 posts 85.8, 67.7, 74.8 and 80.0. Read the columns rather than the summary: Hy-SimulMT wins FLEURS, ACL6060 and RealSI-Sentence among open systems, loses CoVoST2 to Qwen by 6.6 points, and loses FLEURS to Gemini by 2.1. The card's own sentence — higher than the other open-source models on FLEURS, ACL6060 and RealSI-Sentence — is carefully scoped, and the missing benchmark is the one where it does worst. Every number in that table was produced by Tencent. No independent lab has run this model.

The human comparison deserves a closer read than the headline

Tencent also commissioned a matched comparison against professional interpreters: eighteen Chinese and English recordings published online from September 2026 onward, each about six minutes, spanning ten domains from software and medicine to law and agriculture, interpreted by six professionals — three holding CATTI Level 1 certificates and three holding Level 2. For each certification level, the best of the three interpreters was used as the comparison point. The conclusion is that the model reaches quality on par with expert human simultaneous interpreters while often responding faster.

That is a striking result, and it is also a small one, which the card says plainly in a footnote: the findings apply only to those samples, those interpreters and that setting, and do not constitute an assessment of speech quality or voice similarity. Two methodological details are worth holding onto. Using the highest-scoring interpreter at each level sets a demanding bar for quality but an uneven one for latency, since the comparison is against a person rather than against the median professional. And the test measures information delivery and timing, not the thing most listeners notice first in machine speech, which is whether it sounds like a person.

What the repository does not answer

Several questions a buyer would ask first have no published answer yet. There is no hosted API and no price: unlike the Hy-MT2 family, which Tencent lists on Tencent Cloud at published per-token rates, Hy-SimulMT is weights-only plus a mobile app, and Hugging Face's inference-provider panel lists it as not deployed by any provider. There is no technical report and no benchmark outside Tencent's own evaluation. The card states that speech output and voice cloning require a configured TTS service, so the "speech-to-speech" framing describes a system assembled around the model, not one checkpoint that emits audio. llama.cpp support is explicitly not merged upstream — the card recommends waiting for a release that supports the model, even though GGUF builds are published, and the sample environment variable references a checkpoint named hymt-A3b-tgt_only-1019-Q8_0.gguf, an artifact name that does not match the public model IDs.

The language metadata is the most confusing part. The repository tags 36 languages, from Chinese and English through French, Arabic, Khmer, Tibetan, Kazakh and Uyghur. The prose documents bidirectional speech translation between Chinese and English only. Both statements can be true if the extra languages belong to the text backbone inherited from Hy-MT2 rather than to the speech path, but the card does not say so, and anyone planning a Japanese or Arabic deployment should treat that as unconfirmed until Tencent documents it.

Running it yourself, and what that costs in hardware

The documented path is vLLM 0.31 or later with HYV3ForCausalLM support, plus the simulst package. A single command provisions the environment and starts both the model server and a web/API front end on ports 8800 and 8080 respectively; the front end accepts audio uploads, a microphone, or a WebSocket stream, and writes translated audio to disk when TTS is configured. There is also a llama.cpp bridge that runs the web layer on port 8094 against an existing llama.cpp server on 8080, with a MODEL_TYPE switch for translation-only versus source-plus-translation checkpoints. Memory is the practical gate: 30.4 billion parameters in BF16 is roughly 61 GB of weights before KV cache and activations, so a single 80 GB accelerator is the realistic floor and the FP8 build is what makes a smaller card plausible. Those are our figures from the parameter count, not Tencent's minimum specification, which the card does not state.

For teams whose meeting pipeline does more than translate — summarising the transcript, extracting action items, drafting the follow-up — the interpreter is one stage among several, and those later stages are ordinary API calls across whichever vendors happen to be best that month. That is the layer OrcaRouter covers: one API for 200+ models with provider list prices passed through at 0% markup, automatic failover when a provider stumbles, and a routing DSL for composing stages. Hy-SimulMT itself is not on our catalogue and we do not host it; the point is that the surrounding work does not need a separate contract per vendor.

What would change this assessment

Three things are worth watching, in order of how much they would settle. An independent evaluation — the IWSLT circuit publishes StreamLAAL numbers for open systems every year, and a submission using these weights would be the first number not produced by the model's owner. A hosted endpoint with a price, which would tell you whether Tencent intends this as infrastructure or as a feature of its own app. And the llama.cpp merge, which decides whether the GGUF builds are a real local option or a placeholder. Until then, the honest summary is that Tencent has published the most capable open simultaneous speech translator we have seen, with strong vendor numbers, a permissive licence, and no outside confirmation of any of it — and that the timing, one day before an APEC partnership announcement, suggests the confirmation Tencent cares about is a volunteer in Shenzhen holding a phone, not a benchmark leaderboard.