Hy-SimulMT-30B-A3B vs Tencent Hy-MT2-1.8B: A Meeting Room Against a Phone

Tencent now publishes two translation models that share a licence, a lab and almost nothing else. Tencent Hy-MT2-1.8B is the smallest member of the Hy-MT2 text family, open-sourced on 21 May 2026, a dense 1.8-billion-parameter model that compresses to 440 MB at 1.25-bit and runs offline on phone silicon. Hy-SimulMT-30B-A3B is a new line entirely, published as weights on 22 October 2026: 30.4 billion parameters, roughly 3 billion active, built to listen to Chinese or English speech and translate it while the sentence is still being spoken. One fits in a pocket and has been downloaded more than 26,000 times. The other needs an 80 GB accelerator, and four days after release it had zero downloads.

The natural reading of that gap is that the big model is the better one and nobody has noticed yet. The more useful reading is that they answer different questions. Hy-MT2-1.8B answers "how do I translate text on a device with no network?" Hy-SimulMT-30B-A3B answers "how do I translate a conversation without waiting for the speaker to finish?" Tencent shipped both within five months, which tells you it considers those separate markets rather than points on a size curve.

Not a size variant: a different organ

It is easy to assume Hy-SimulMT-30B-A3B is the 30B tier of Hy-MT2, the way the 7B and the 30B-A3B MoE sit above the 1.8B. It is not. Hy-MT2 is a text-only family with a shared tokenizer, a shared instruction format and a shared benchmark suite called IFMTBench; its three sizes are trained to do the same job at different quality and cost points. Hy-SimulMT is a separate architecture with an audio encoder bolted to a mixture-of-experts backbone: 24 encoder layers at width 1,024 projecting into a 2,048-wide model, 128 experts of which eight are routed plus one shared per token, and a maximum input of thirty seconds of 16 kHz audio per window.

That distinction matters for anyone planning a migration. You cannot swap a Hy-MT2-1.8B call for a Hy-SimulMT call by changing a model ID, because one takes text and the other takes audio, and the speech model's output is a translation of what it heard rather than of what you typed. The overlap is conceptual — both translate Chinese and English — not operational.

What the 1.8B is genuinely good at

The small model's case is made by three numbers. The first is 440 MB, which is the size of the 1.25-bit quantisation produced with Tencent's AngelSlim tooling and the reason the model runs on Apple, Qualcomm and MediaTek phone chips; Tencent claims it is about 1.5 times faster than the previous Hy-MT1.5 generation at 4-bit on an A15. The second is 90.00 on FLORES-200 English-to-X, which Tencent presents as roughly 89.9% of Gemini 3.1 Pro's score — a vendor comparison, unreproduced, but a striking one for a model that fits on a handset. The third is the download count: more than 26,000 on the main repository and well over a thousand likes, against 3,132 downloads for the 30B-A3B flagship. The small one is the one people actually use.

Its documented limits are equally clear. The working context is 8,192 tokens with a 4,096-token output cap, even though the configuration advertises 262,144 positions; Tencent's own inference guidance is the number to trust. It supports 33 languages plus five Chinese minority languages and dialects. And its weakest published result is instruction-following: 69.36% overall on IFMTBench, against 85.7% for the family's 30B-A3B sibling. If your job is translating a clean sentence, the 1.8B is competitive. If your job is honouring a glossary, preserving JSON structure and matching a register, it is the least reliable model in its own family, and the gap is not small.

What the 30B buys, and what it does not

Hy-SimulMT-30B-A3B's purchase is latency in a conversational setting. Tencent reports approximately three seconds of StreamLAAL on long-form RealSI audio, described as the lowest among the systems compared, and 84.7 on RealSI-Sentence, the highest in its table. Those are speech-translation metrics from the vendor's own evaluation, with no independent reproduction. The model also supports real-time voice cloning, so the translated speech can keep the speaker's vocal characteristics — but only if you connect a TTS service, which the deployment toolkit expects you to configure. The weights do not synthesise audio on their own.

What it does not buy is language coverage. The card documents bidirectional speech translation between Chinese and English. The repository tags 36 languages, which most plausibly reflects the text backbone rather than the speech path, but Tencent does not explain the discrepancy. A team that needs Japanese-to-English speech translation has no documented path with this model, while the 1.8B handles Japanese as text today. The big model is not a superset of the small one; it is a specialist with a narrower lane and a much deeper cut.

Where the 440 MB actually lives

The interesting thing about the 1.8B is not its benchmark score but its deployment envelope. A model that runs offline on a phone changes what a translation product can promise: no round trip, no per-token bill, no data leaving the device, no dependence on connectivity in a basement conference room or on a plane. Tencent ships it inside the Hy Translate app, which launched in Chinese app stores on 24 September 2026 and offers downloadable offline models alongside text, photo, voice and turn-based conversation modes. The app covers 12 countries and regions and supports 33 languages plus five minority languages and dialects.

That app is also the clearest evidence of where Hy-SimulMT fits. In September, the conversation mode was turn-based — two people alternate speaking and read the translation. A simultaneous model is what turns that into an interpreter you can wear. Tencent's card says Hy-SimulMT can be tried in the Hy Translate app, and the timing of the weights, one day before Tencent announced its APEC 2026 partnership, suggests the intended showcase is a live event rather than a developer API. That is inference from public dates, not a stated plan.

Cost and deployment, point by point

• Parameters — Hy-SimulMT-30B-A3B: 30.4B total, roughly 3B active, sparse MoE plus audio encoder; Hy-MT2-1.8B: 1.8B dense, text only.

• Footprint — Hy-SimulMT: ~61 GB in BF16 before KV cache, FP8 builds published; Hy-MT2-1.8B: 440 MB at 1.25-bit, with FP8 and GGUF builds also available.

• Hardware — Hy-SimulMT: one 80 GB accelerator realistically; Hy-MT2-1.8B: a phone, a laptop CPU, or a single small GPU.

• Context — Hy-SimulMT: thirty-second audio windows; Hy-MT2-1.8B: 8,192 tokens in, 4,096 out.

• Languages — Hy-SimulMT: Chinese and English documented for speech; Hy-MT2-1.8B: 33 languages plus five minority languages and dialects.

• Price — Hy-SimulMT: no hosted API or published rate; Hy-MT2-1.8B: about $0.044 per million input tokens and $0.177 per million output tokens on Tencent's hosted API since August 2026.

• Licence — both Apache-2.0, both ungated, both fine-tunable and redistributable.

Neither model is on OrcaRouter's catalogue and we host neither. The routing question they raise is different from the usual one: if the 1.8B runs on the device and the 30B runs on your own GPU, the calls that need a router are the ones around them — the document ingestion, the transcript cleanup, the summariser that turns a translated meeting into minutes. Those are ordinary cloud model calls, and putting them behind one API for 200+ models at 0% markup over provider list prices keeps the vendor count down while the translation stage stays self-hosted. Failover matters most precisely when the model doing the critical work is one you cannot swap quickly.

Two products, not two tiers

If you need translation on a device, offline, at zero marginal cost, and your content is text, Tencent Hy-MT2-1.8B is the model — it is small, permissively licensed, widely deployed, and honest about its instruction-following weakness. If you need to translate a live conversation between Chinese and English and a three-second lag is acceptable, Hy-SimulMT-30B-A3B is the only Tencent model that does it, and the fact that it is thirty times larger is the price of the audio encoder and the latency budget, not a quality upgrade you can bank elsewhere.

The trap is assuming the newer, bigger release supersedes the older, smaller one. It does not. One of these models has 26,000 downloads because it solves a problem millions of people have; the other has zero because it solves a problem that, so far, mostly belongs to Tencent's own app and to the APEC volunteers who will be using it in Shenzhen this November.