Tencent Hy-SimulMT-30B-A3B Explained: Open-Source Chinese-English Simultaneous Translation at About 3 Seconds of Latency

Quick Answer

Tencent's open Chinese-English simultaneous translation model: about 3 seconds of latency, a 30B MoE activating only 3B per token, Apache-2.0.

On October 26, 2026, Tencent Hunyuan released and open-sourced Hy-SimulMT: a bidirectional Chinese-English speech-to-speech simultaneous translation model — the speaker talks, and the translation comes out as speech and text at the same time, with real-time voice cloning that keeps the speaker's own vocal characteristics. The model collection is on Hugging Face (this article's scanned term is the FP8 build, tencent/Hy-SimulMT-30B-A3B-FP8), ModelScope, and GitHub, with a Hy Translate App for trial. Three things to get straight first: all latency and quality figures are vendor-reported (preliminary) with no independent reproduction yet; it covers Chinese-English only, not the 33 languages of the Hy-MT2 text family; and no hosted Hugging Face API is available — deployment means self-hosting.\n\nDisclaimer: this site is not affiliated with Tencent. This article is product-technology reporting and analysis, not procurement, deployment, or investment advice. Latency and quality figures are vendor-reported and marked preliminary; deployment and license details follow official pages and may change.

1. What it is: Chinese-English speech interpretation that starts before the speaker finishes, open-sourced three days ago

Interpretation is not translation: ordinary translation waits for a sentence to finish, interpretation has to start while the speaker is still talking, and latency itself is part of quality. Hy-SimulMT is built for that setting: it continuously receives the speaker's audio and translates as it listens, producing translated text and speech together, with real-time voice cloning so the output keeps the speaker's own voice instead of swapping in a standard announcer voice. The target settings are named concretely: meetings, speeches, and interviews — live communication.\n\nThe timeline: released and open-sourced by Tencent Hunyuan on October 26, 2026, with the Hugging Face model pages, ModelScope, and the GitHub repository going live together, plus a Hy Translate App for trial. Written on October 29, the term is three days old — no staleness question. One boundary to draw early, because the names invite confusion: the 30B-A3B in this name matches the Hy-MT2-30B-A3B from May 2026 — a 30B-total, about-3B-active MoE — but Hy-MT2 is a 33-language text translation family, while Hy-SimulMT covers Chinese-English speech only. Do not size this model by the text family's language list.

2. Latency and quality: what the about-3-seconds figure measures, and where it beats humans

The latency metric is StreamLAAL (streaming average lagging): roughly, how long after speech arrives the translation is heard. In the vendor's RealSI long-form audio evaluation the figure is about 3 seconds, the lowest among the systems shown. A more concrete comparison puts it against six certified simultaneous interpreters (three CATTI Level 1, three Level 2) across ten audio domains — software AI, medicine, manufacturing and energy, business management, and others — at 3.3 seconds, against 4.5 seconds for Level 1 and 4.7 seconds for Level 2 interpreters. Recomputed here: about 27% faster than Level 1 and about 30% faster than Level 2. On quality, the vendor claims translation on par with human simultaneous interpreters, with GEMBA scores above the compared open-source models and four of the five commercial APIs.\n\nAll of this must stay labeled: vendor-reported (preliminary), no independent reproduction; and the two latency figures (about 3 seconds and 3.3 seconds) come from different evaluation cuts — do not merge them into a single number when quoting. One third-party caveat deserves to stay verbatim in spirit: the study does not establish equivalent performance across medical, legal, technical, or noisy conversational speech — the home ground is clean live settings like talks and interviews, not high-stakes professional domains or noisy rooms.

3. The MoE and the six variants: 30B of knowledge, 3B of compute

The architecture is a mixture of experts: 30B total parameters, but only about 3B active per token — recomputed here as 10% of parameters per step. That split is what makes a 30B-class translator viable in real time: the compute bill follows the 3B active figure, while the memory bill follows the full 30B of weights. Keep the two budgets separate.\n\nThe six variants split on two axes: output mode — a plain translated-text build and an Interleave build that interleaves source text with translation for side-by-side checking — and precision — full-precision, FP8, and GGUF builds. The FP8 build, this article's scanned term, is the deployment story: on weight math, FP8 at 1 byte per parameter puts 30B at about 30GB against about 60GB for BF16 — moving single-GPU serving from “two high-end cards” to “one large-VRAM card” territory (estimates, excluding runtime overhead; follow official configs in practice). The GGUF builds target llama.cpp, with one catch: official llama.cpp support has not merged yet, so use the project's integration path or a compatible build — do not schedule it as download-and-run. The repository also ships SimulInfer (web and WebSocket inference, optional TTS integration; the deploy script starts vLLM and the web/API service on port 8080) and SimulTrain training tools; speech output and voice cloning need a separately configured TTS service, which is not inside the weights.

4. Boundaries and selection: Chinese-English only, and high-stakes settings are outside the proven range

Three tiers. Tier one, direct fits: bilingual meetings, talks streamed live, interviews — latency already below the professional average, Apache-2.0 is commercially friendly, and FP8 serving fits a single card, so pilot it in a small live setting first. Tier two, usable with spot checks: real meetings with accents, fast speech, and domain terminology — the vendor evaluation covered ten domains, but your domain mix may differ from the evaluation mix, so run your own recordings through StreamLAAL and sampled translation checks before committing. Tier three, not recommended yet: languages beyond Chinese-English (out of scope), high-stakes professional settings like medical and legal work, and noisy conversational environments — equivalence there is explicitly unestablished, and where mistranslation is expensive, human interpreters should stay primary with the model as an assistive record.\n\nTwo deployment notes. First, access: no hosted Hugging Face API is available, so plan on self-hosting SimulInfer, including time for service setup and TTS configuration. Second, voice cloning stays at the feature level here: it preserves the speaker's voice for translation, and public deployments need speaker authorization and translation labeling per the official terms — this article gives no tutorial for cloning a specific person's voice. Review scores (preliminary, from public sources, not our own testing): latency-lead credibility medium (vendor-reported, consistent across sources, but no independent reproduction, and the two latency figures use different cuts); deployment feasibility credibility high (variants, license, and service components are specific and checkable, and the weight math recomputes cleanly here); cross-domain generalization credibility low (equivalence in high-stakes and noisy settings is explicitly unestablished).

5. Sources and disclaimer

Sources (verified 2026-10-29):\n1) https://huggingface.co/tencent/Hy-SimulMT-30B-A3B — official model page: bidirectional Chinese-English speech-to-speech simultaneous translation, translating while listening, real-time voice cloning, RealSI StreamLAAL about 3 seconds, SimulInfer and SimulTrain, Hy Translate App;\n2) https://huggingface.co/tencent/Hy-SimulMT-30B-A3B-FP8 — this article's scanned term: the FP8 build;\n3) https://www.kucoin.com/news/flash/tencent-s-hy-simulmt-achieves-sub-3-second-delay-in-real-time-translation — October 26, 2026 release, 3.3 seconds vs 4.5 for Level 1 and 4.7 for Level 2, ten audio domains, six certified interpreters, quality on par with human interpreters;\n4) https://alphasignal.ai/news/tencent-s-hy-simulmt-translates-chinese-and-english-before-speakers-finish — six variants, 30B MoE with about 3B active per token, FP8 single-GPU serving, llama.cpp support not yet merged, SimulInfer on port 8080, equivalence in medical, legal, and noisy settings unestablished;\n5) https://huggingface.co/tencent/Hy-MT2-30B-A3B-FP8 — Hy-MT2 family background: 30B-A3B MoE naming and the 33-language text translation positioning (used to draw SimulMT's Chinese-English boundary).\n\nDisclaimer: this site is not affiliated with Tencent. This article is product-technology reporting and analysis, not procurement, deployment, or investment advice. Latency and quality figures are vendor-reported, marked preliminary, and not yet independently reproduced; deployment and license details follow official pages and may change.

Latency comparison: Hy-SimulMT vs professional interpreters (vendor-reported, preliminary)

Subject	Latency (StreamLAAL)	Quality basis	Languages and access
Hy-SimulMT-30B-A3B	About 3 seconds (3.3 seconds in the interpreter comparison)	On par with human interpreters; GEMBA above 4 of 5 commercial APIs	Chinese-English only; Apache-2.0 open, self-hosted
CATTI Level 1 interpreters	4.5 seconds	Human baseline	Many languages; human service
CATTI Level 2 interpreters	4.7 seconds	Human baseline	Many languages; human service

Simultaneous interpretation latency in seconds (lower is better; vendor-reported, preliminary)

★ Hy-SimulMT3.3 seconds

CATTI Level 14.5 seconds

CATTI Level 24.7 seconds

数据来源：Sources: Tencent's reported comparison figures (consistent across multiple reports, 2026-10-29); no independent reproduction yet, for order-of-magnitude reference only

Frequently Asked Questions

How is Hy-SimulMT's 3-second latency measured?

The metric is StreamLAAL (streaming average lagging): the average wait between speech arriving and translation being heard. The vendor's RealSI long-form evaluation gives about 3 seconds; a separate comparison against six certified interpreters gives 3.3 seconds. The two figures come from different evaluation cuts — do not merge them into one number when quoting. All vendor-reported, marked preliminary, with no independent reproduction.

Which languages does Hy-SimulMT support?

Chinese-English only, in both directions: Chinese speech to English and English speech to Chinese, translating as the speaker talks. That is its biggest difference from the Hy-MT2 text family (33 languages); scenarios needing a third language are out of scope.

What does local Hy-SimulMT deployment need?

The FP8 weights are estimated at about 30GB (about 60GB for BF16), aimed at single large-VRAM GPU serving. The repository provides SimulInfer (web and WebSocket, port 8080); speech output and voice cloning need a separately configured TTS service. GGUF builds target llama.cpp, but official support has not merged yet, so use the project's integration path. No hosted Hugging Face API is available.

Could the voice cloning feature be misused?

Its product purpose is keeping the speaker's own voice in translation, not impersonation; public deployments need speaker authorization and translation labeling per the official terms. This article stays at the feature level and gives no tutorial for cloning a specific person's voice.