Hy-SimulMT-30B-A3B vs Helsinki-NLP OPUS-MT: A Live Interpreter Against a Folder of Specialists

Picture the same conference talk handled two ways. In the first, a microphone feeds an automatic speech recogniser, the recognised sentence goes to a translation model, the translation goes to a speech synthesiser, and the listener hears English a few seconds after the speaker said the Mandarin — with every name, number and acronym having survived three separate chances to be mangled. In the second, a single model listens to the audio and emits the translation as the sentence unfolds. Hy-SimulMT-30B-A3B, released as weights by Tencent on 22 October 2026, is built for the second. Helsinki-NLP OPUS-MT is the most widely deployed open translation system in the world and is built for a narrower version of the first: it translates text, one sentence at a time, one language direction per checkpoint, and it has been downloaded hundreds of thousands of times because it does that job on a CPU for free.

These two are not rivals in the ordinary sense. One is a 30.4-billion-parameter speech model that needs an accelerator; the other is a collection of more than a thousand bilingual checkpoints of roughly 70–80 million parameters each, most of them around 300 MB, which run comfortably on a laptop. The useful comparison is not which one translates better. It is where the boundary sits between a translation component and a translation system, and how much of that boundary you want to own.

What each one actually is

OPUS-MT is maintained by the University of Helsinki's NLP group and built on the MarianMT encoder-decoder architecture. It is not a model but a catalogue: separate checkpoints for Chinese-to-English, English-to-Chinese, English-to-German and several hundred other directions, each trained on parallel corpora from the OPUS collection, each downloadable in isolation. The Chinese-English checkpoint alone has been downloaded nearly half a million times; the reverse direction has seen about 73,000. There are multilingual variants that fold many source languages into one English-target model, but the common case is a narrow specialist that knows exactly one job.

The architecture explains both its virtue and its ceiling. A 77-million-parameter encoder-decoder trained only on parallel text has no room for world knowledge, no instruction interface, and no mechanism for streaming. Feed it a sentence and it returns a sentence. Feed it a paragraph and it translates the paragraph as a sequence of sentences with no shared context, and published engineering notes report a tendency to silently drop content on long inputs. It cannot be told to keep a glossary term, preserve a JSON key, or match a formal register — you handle those with preprocessing and post-editing, or with constrained decoding if you are prepared to build it.

Hy-SimulMT-30B-A3B is a different species of object. Its configuration describes a 48-layer sparse mixture-of-experts model with 128 experts, eight routed plus one shared per token, and a 24-layer audio encoder feeding a 2,048-wide backbone. The name's "A3B" is not expanded in the card, but the config arithmetic lands at roughly 3 billion active parameters against 30.4 billion total. It takes 16 kHz audio in thirty-second windows and produces translated text; the Interleave variant also emits the source transcription. Text-to-speech and voice cloning are handled by a separate service that the deployment toolkit connects to, which means the "speech-to-speech" product is an assembly, not a single file.

The latency argument is really an error-propagation argument

Latency is the obvious difference and the least interesting one. A cascade can be tuned to be fast; a good streaming ASR with partial hypotheses plus a small MT model can produce a usable subtitle stream well under a second behind the speaker. The reason simultaneous models exist is that speed in a cascade is bought by giving each stage less context, and each stage then makes errors the next stage cannot see. The recogniser mishears a proper noun, the translator has no way to know it was a name, and the synthesiser pronounces the error confidently. The listener has no way to tell which stage failed.

A single model trained end-to-end on the whole task can, in principle, defer a decision until the audio resolves the ambiguity, and it can carry the speaker's voice through if voice cloning is enabled. Tencent's claim is precisely about this trade: a StreamLAAL of approximately three seconds on long-form RealSI audio, described as the lowest among the systems compared, with a GEMBA quality score above the open-source models in its table and four of five commercial APIs. Whether three seconds is good depends entirely on the setting — it is a long time in a negotiation and a short time in a lecture — but it is measured on the streaming task, which is the task OPUS-MT cannot attempt without a cascade built around it.

Coverage, counted honestly

This is where the folder wins, decisively, and it is the part most comparisons blur. OPUS-MT's thousand-plus checkpoints span roughly a hundred languages across various directions, including low-resource pairs that no large multilingual model has ever seen. Hy-SimulMT documents bidirectional speech translation between Chinese and English. The repository's metadata tags 36 languages — French, Arabic, Khmer, Tibetan, Kazakh and Uyghur among them — but the prose describes only the Chinese-English pair, and nothing in the card explains the discrepancy. The most likely reading is that the tags describe the text backbone inherited from Tencent's Hy-MT2 family rather than the speech capability, but that is an interpretation, not a published fact.

So the honest coverage line is: thirty-six languages claimed in metadata, two languages documented for speech, against a catalogue of a hundred-odd languages in text. If your requirement is Mandarin to English in a live meeting, Hy-SimulMT is aimed exactly at you. If your requirement is Polish to Finnish on a static document, OPUS-MT has a checkpoint for it and Hy-SimulMT does not.

Quality evidence: one side has independent numbers, the other has a vendor table

The independent evidence runs in OPUS-MT's favour, on the narrow question of text translation quality. A published study on Fassa Ladin translation measured an OPUS-MT model at 55.61 BLEU Italian-to-English against 52.93 for NLLB-200 and 44.18 for M2M-100, and 44.35 English-to-Italian against 40.13 and 32.40. The lesson of that study is that a model trained exclusively on one direction can still beat far larger multilingual systems on that direction. It says nothing about speech, and it does not involve Hy-SimulMT, because no independent lab has evaluated Hy-SimulMT at all.

On the other side, Tencent reports 90.3 FLEURS, 75.5 CoVoST2, 81.2 ACL6060 and 84.7 RealSI-Sentence for Hy-SimulMT, all from its own evaluation. Those are speech-translation benchmarks, so they are not comparable to the BLEU figures above, and they carry no third-party confirmation. The one internal signal worth noting is that the model loses CoVoST2 to Qwen3.8-LiveTranslate by 6.6 points while winning the other three columns among open systems — a reminder that a single number would have hidden a real weakness.

Cost and deployment, side by side

• Parameters — Hy-SimulMT-30B-A3B: 30.4B total, roughly 3B active, sparse MoE; Helsinki-NLP OPUS-MT: about 70–80M per checkpoint, dense.

• Footprint — Hy-SimulMT: roughly 61 GB of BF16 weights before KV cache, so an 80 GB accelerator is the practical floor; FP8 builds are published. OPUS-MT: about 300 MB per checkpoint, CPU-capable.

• Latency model — Hy-SimulMT: streaming, approximately 3 s StreamLAAL on long-form audio. OPUS-MT: sentence-at-a-time; streaming requires an ASR and a segmentation policy you build.

• Languages — Hy-SimulMT: Chinese and English documented for speech, 36 tagged in metadata. OPUS-MT: roughly a hundred languages across more than a thousand directional checkpoints.

• Licence — Hy-SimulMT: Apache-2.0 for Tencent's weights and code, with third-party components under their own terms. OPUS-MT: per checkpoint and not uniform — the Chinese-to-English model is CC-BY-4.0, the English-to-Chinese model is Apache-2.0, English-to-German is CC-BY-4.0. Check the specific card before shipping.

• Hosted API — Hy-SimulMT: none published. OPUS-MT: none; weights only.

That last pair matters for procurement. Neither model is on OrcaRouter's catalogue and we do not host either — you self-host the weights or use the vendor's own channel. Where routing does earn its place is in the rest of the pipeline: the transcript summariser, the glossary checker, the post-editor and the chat assistant that sit around the translation stage are ordinary model calls, and running them through one API for 200+ models at 0% markup over provider list prices avoids a separate contract per vendor. Translation is rarely the only thing a meeting product does.

Which one belongs in your stack

Choose OPUS-MT when the input is already text, the direction is fixed, the volume is high, and the budget is zero. Legal document pipelines, subtitle files, e-commerce catalogues and offline browser translation are all still served better by a 300 MB specialist than by a 30-billion-parameter model, and the per-checkpoint licence is the only thing you need to audit. Choose Hy-SimulMT when the input is audio and the listener is waiting: live meetings, lectures, interviews, broadcast interpretation, any setting where a three-second lag and a preserved voice are worth an accelerator and a TTS service.

The mistake to avoid is treating this as a quality contest. OPUS-MT does not lose to Hy-SimulMT on text translation; it loses on a task it was never built for. And Hy-SimulMT does not replace a thousand careful bilingual models with one clever one — it replaces a cascade you would otherwise have to assemble, tune and debug yourself. Those are different purchases, and only one of them comes in a folder.