Hy-SimulMT-30B-A3B vs Tencent Hy-MT2-7B: The Price You Can Quote vs the One You Can't

Ask what an hour of Chinese-to-English interpretation costs and the two models in this comparison give you completely different kinds of answer. Tencent Hy-MT2-7B, open-sourced on 21 May 2026, has a rate card: $0.074 per million input tokens and $0.295 per million output tokens on Tencent's hosted API. Hy-SimulMT-30B-A3B, published as weights on 22 October 2026, has no published price at all — no endpoint, no per-token rate, no public cloud listing. You can budget the 7B down to the cent. For the speech model you have to budget a GPU, and the number you arrive at depends entirely on what you already own.

That difference is more than a procurement detail. It reflects two different products with two different cost structures. The 7B is a text translator priced like a utility: you pay for tokens, you scale with volume, and the marginal cost of one more document is nearly zero. Hy-SimulMT is a real-time system priced like infrastructure: you pay for an accelerator by the hour whether or not anyone is speaking, and the marginal cost of one more meeting is close to nothing only because the machine is already running. Comparing them by parameter count misses the point. Comparing them by what an hour of work actually costs is the whole argument.

What the 7B is, and what it costs per token

Tencent Hy-MT2-7B is the middle child of the Hy-MT2 text family, a dense 7-billion-parameter decoder trained for translation across 33 languages plus five Chinese minority languages and dialects. It has an 8,192-token working context, supports structured output through response_format, and does not support function calling. In BF16 it needs roughly 15 GB of accelerator memory, about 8 GB in FP8, and a few gigabytes in GGUF form — a single mid-range GPU, or a CPU if you are patient.

Its vendor-reported quality sits close to the frontier on the benchmarks Tencent chose. On FLORES-200 it scores 86.89 XCOMET-XXL, which Tencent presents as about 97.9% of Gemini 3.1 Pro in thinking mode. On DomainMTBench it reports 94.92 and 92.79 on two splits; on IFMTBench, the instruction-following suite Tencent built for this family, 83.14 overall. None of these has been independently reproduced. They are the vendor's numbers, and they describe a model that is good at controlled translation — glossaries, register, structured output — which is exactly the gap that separates the 7B from the 1.8B below it.

At the hosted rate, a million tokens of input costs seven cents and a million tokens of output costs under thirty. For text work that is close to free. The interesting question is what happens when the input is not text.

The arithmetic of an hour of speech, both ways

An hour of continuous speech runs to roughly 9,000 English words, which is somewhere around 12,000 tokens. Put that through Tencent Hy-MT2-7B at list price and the translation itself costs about half a cent: roughly $0.0009 for the input and $0.0035 for the output, so call it $0.0044. Even if you are off by a factor of two on the token estimate, the number is still fractions of a cent.

But the 7B cannot hear. A text pipeline starts with a recogniser, and a hosted recogniser typically costs a few tenths of a cent per minute, which puts an hour of audio somewhere under a dollar before the translator runs at all. That is the first honest correction to the "7B is cheaper" story: in a cascade, the translator is not the expensive stage. The recogniser is. The second correction is latency — a cascade that waits for a sentence boundary, transcribes, translates and displays will trail the speaker by more than a simultaneous model does, and no amount of cheap tokens fixes that.

Hy-SimulMT-30B-A3B inverts the cost structure. There is no per-token bill because there is no hosted service; you rent or own the hardware. The BF16 weights are about 61 GB before KV cache, so an 80 GB accelerator is the realistic floor, and a single card rents for anywhere from roughly $2 an hour at a specialist GPU cloud to more than $10 on a hyperscaler's on-demand list. At the low end, an hour of meeting costs about $2 in compute; at the high end, an hour costs an hour's rental. Both figures assume the machine is dedicated to that meeting, which is the worst case — a server handling several concurrent streams divides the cost. Running the FP8 build on a smaller card moves the floor down again.

So the comparison is roughly $0.30–$1.00 per meeting-hour for the cascade, dominated by recognition, against $2–$12 per meeting-hour for the simultaneous model, dominated by the GPU. That is a real difference of one to two orders of magnitude, and it is the correct frame: you are not paying for better translation, you are paying for three seconds of lag instead of fifteen. Teams that need the lag will pay it. Teams that do not are buying an accelerator to solve a problem they do not have.

Quality: two different rulers

There is no honest way to put Tencent Hy-MT2-7B's 86.89 FLORES-200 score next to Hy-SimulMT's 84.7 on RealSI-Sentence. One measures text translation against a reference; the other measures speech translation on a benchmark built for streaming audio. They share a number format and nothing else. What can be compared is the shape of the evidence. The 7B has a full published table across FLORES-200, WMT25, DomainMTBench and IFMTBench, all vendor-run. The speech model has a four-column table across FLEURS, CoVoST2, ACL6060 and RealSI-Sentence, also vendor-run, plus a human comparison against six professional interpreters that Tencent itself describes as small-scale and limited to the samples used. Neither has independent verification.

The one place the two lines genuinely meet is translation quality on Chinese and English. The 7B is the stronger text translator by every measure Tencent publishes, including instruction-following, where its 83.14 on IFMTBench is fourteen points above the 1.8B. The speech model's advantage is not quality but timing: it produces a usable translation while the sentence is still in progress, and it can carry the speaker's voice through a TTS service if one is configured. If your audio is already transcribed and your latency budget is measured in minutes, the 7B is the better translator. If your latency budget is measured in seconds, the 7B is not in the running regardless of its score.

Deployment and cost, side by side

• Parameters — Hy-SimulMT-30B-A3B: 30.4B total, roughly 3B active, sparse MoE with an audio encoder; Tencent Hy-MT2-7B: 7B dense, text only.

• Memory — Hy-SimulMT: ~61 GB BF16, FP8 build published; Hy-MT2-7B: ~15 GB BF16, ~8 GB FP8, a few GB in GGUF.

• Context — Hy-SimulMT: thirty-second audio windows; Hy-MT2-7B: 8,192 tokens.

• Latency — Hy-SimulMT: approximately 3 s StreamLAAL, vendor-reported; Hy-MT2-7B: text in, text out, with a cascade's ASR delay in front of it.

• Languages — Hy-SimulMT: Chinese and English documented for speech; Hy-MT2-7B: 33 languages plus five minority languages and dialects.

• Price — Hy-SimulMT: none published; Hy-MT2-7B: $0.074 per million input tokens, $0.295 per million output tokens.

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

Neither model is on OrcaRouter's catalogue and we host neither. The pricing difference is still the reason to care about routing here. Tencent prices its translation APIs aggressively and changes those prices: on 12 June 2026 it cut Hy-MT2-Pro to ¥0.5 per million input and ¥2 per million output tokens, a 66.7% cut on input and 55.6% on output. OrcaRouter passes provider list prices through at 0% markup, so a cut like that is live on our side the same day rather than at the next contract renewal, and automatic failover means one provider's outage does not take a translation pipeline down. For a workload whose cost is dominated by per-token rates, the difference between a 0% and a 20% markup is the whole margin.

Which one your latency budget picks

Start with the budget. If the answer is "seconds", nothing about the 7B's price matters — a cascade cannot produce a simultaneous feed without a streaming recogniser and a segmentation policy you build and tune, and even then it will trail. Hy-SimulMT-30B-A3B is the only one of these two built for that constraint, its Apache-2.0 licence means you can test the claim on your own audio, and the cost is a GPU by the hour. Budget $2–$12 per meeting-hour as a starting assumption and check it against your own utilisation.

If the answer is "minutes", the 7B wins on every axis that matters: cost, memory, language coverage, documented instruction-following, and the existence of a price you can put in a spreadsheet. Pay the recogniser, pay half a cent for the translation, and spend the difference on the review step. Most translation work in most companies is still this second kind, which is why a seven-billion-parameter model with a rate card is likely to outlive a thirty-billion-parameter model whose most interesting feature is that it does not have one yet.