Tencent Hy-MT2-7B vs Tencent Hy-MT2-30B-A3B: The Dense Sweet Spot or the MoE Flagship at the Same Price

Here is the surprise at the heart of this matchup: Tencent Hy-MT2-7B, a dense 7B model, and Tencent Hy-MT2-30B-A3B, the mixture-of-experts flagship with 30 billion total parameters and roughly 3 billion active, both became callable through hosted APIs this week — the 7B around August 19, 2026, the 30B-A3B a day later, both served by Tencent Cloud — at the exact same price: $0.074 per million input tokens and $0.295 per million output. The family was open-sourced together on May 21, 2026, so neither model is new; the identical API price is what makes this comparison genuinely strange. Usually the bigger model costs more and you weigh the premium against the gain. Here the flagship costs nothing extra per token, and the decision collapses into a single question: what, if anything, does the 7B give up by being dense and small?

The same-price surprise

The 30B-A3B's price parity is the MoE bargain doing its job. Its architecture holds 30B of stored knowledge but activates only about 3B per token, so Tencent Cloud can serve it at the same per-token rate as the 7B — same $0.074 / $0.295, same 8,192-token context, same structured-output support, same no-function-calling limitation. On the API, the "professional" model is not more expensive. The catch moves elsewhere: into memory footprint, serving complexity, and latency, where the 7B's denser, simpler design has structural advantages that a per-token price cannot express.

Specs, side by side

• Architecture — dense ~7B vs MoE 30B total / ~3B active.

• API price — $0.074 / $0.295 vs $0.074 / $0.295 per 1M tokens (identical).

• Context — both 8,192 tokens.

• Positioning — balanced sweet spot vs professional-grade flagship.

• FLORES-200 — 7B: 86.89 XCOMET-XXL, ≈97.9% of Gemini 3.1 Pro (Think); 30B-A3B: the family's best general-translation score (figures vendor-reported).

• Deep​Seek / K​imi comparison — both beat DeepSeek-V4-Pro and Kimi K2.6 in fast-thinking mode, vendor-reported.

• Footprint — single A100 / RTX 4090 vs 30B-scale weights (an 18GB-class quantized build on disk and more in VRAM).

• Inference defaults — temp 0.7, top_p 0.6, top_k 20 vs temp 0.7, top_p 1.0, top_k -1.

Where the 30B-A3B actually wins

Tencent positions the 30B-A3B as the professional-grade member of the family, and the evidence it publishes backs that label for a specific kind of text. On domain-heavy material — legal clauses, medical text, technical documentation — its GEMBA line on DomainMTBench is the strongest in the family, reported at around 99% of the Gemini 2.5 Pro baseline, and its general-translation score tops the 7B's on the family's own FLORES curve. That dormant 27B of extra knowledge is exactly the kind of capacity that shows up when a sentence carries dense terminology rather than ordinary prose: a contract's "indemnification shall survive termination" does not tax a 7B model much, but a full M&A document with a glossary does. The 30B-A3B is also the safest bet for Chinese-to-minority-language pairs (Tibetan, Uyghur, Mongolian), where Tencent reports its best numbers.

Where the 7B wins anyway

The 7B's advantages are operational, and they are real. First, self-hosting: the 7B fits on a single A100 or RTX 4090 and has the widest framework coverage — Transformers, vLLM, SGLang, and llama.cpp via GGUF are all exercised. The 30B-A3B, even quantized, wants data-center-scale VRAM and fewer people have put it through production serving stacks. Second, latency and serving simplicity: a dense 7B is easier to keep hot, and its recommended sampling is closer to a standard decode. Third, on the API, the identical price means the 7B costs exactly the same per token as the flagship — so the only reason to take the smaller model is that for clean general text, the quality gap is too thin to notice. The 7B is already roughly 97.9% of Gemini 3.1 Pro (Think) on FLORES-200; the flagship's extra points sit on top of a curve where each one is harder to see in practice.

The gap, honestly measured

Everything in the last two sections is Tencent's own evaluation, and the honest way to read it is that the two models are close enough on ordinary translation that your corpus decides. On general clean text, the 7B's ~97.9%-of-Gemini score means the flagship's margin is measured in small XCOMET fractions — real in a leaderboard, hard to feel in a support ticket. On dense domain text and minority-language pairs, the flagship's reported GEMBA and FLORES leads are exactly where a professional translator (human or model) earns its keep, and where a retranslation is most expensive. There is no independent benchmark for either model at this writing; the one thing both vendor cards agree on is that each size beats DeepSeek-V4-Pro and Kimi K2.6 in the family's fast-thinking mode.

Routing the family

Neither model is on OrcaRouter's catalogue as of this writing, so both are served through Tencent's own cloud and several third-party platforms. The routing lesson is still the practical one. Because the API prices are identical, this is a textbook case for workload routing rather than a single pick: send the high-volume general-translation slice to the 7B and the dense, domain-heavy slice to the 30B-A3B, with automatic failover so a latency spike on the MoE endpoint never stalls the pipeline. That is the pattern OrcaRouter's routing DSL composes across the 200+ models we do carry — cost-weighted dispatch, provider list price passed through at 0% markup — and it fits this family better than it fits most, because the vendor has priced the two tiers to make the split economically neutral.

The verdict

At identical API prices, the default answer is the 7B: same per-token cost, simpler to self-host, and within a whisker on general text. Take the 30B-A3B when the corpus is professionally dense — legal, medical, technical, or minority-language translation — where the flagship's reported domain advantage is worth the larger footprint and the latency. And if your workload is a mix of both, do not pick: route the general slice to the 7B and the hard slice to the flagship. That is not a compromise between the two; it is the only way to get both of them at the price Tencent set.