Tencent Releases Hy-MT2-30B-A3B, a 30B-Parameter Translation Model with 3B Active Parameters

Tencent has released Hy-MT2-30B-A3B, the flagship model in its Hy-MT2 translation family, according to a listing on OpenRouter. The model uses a mixture-of-experts architecture with 30 billion total parameters and 3 billion active parameters per inference pass, denoted by the "A3B" suffix.

What the model does

Hy-MT2-30B-A3B supports 33 language pairs alongside five Chinese dialect and minority-language pairs, according to Tencent. The model is built around five distinct translation workflows:

• Structured translation — for documents with defined formatting
• Delimiter-based translation — for segmented text
• Contextual translation — accounting for surrounding text
• Glossary-based translation — enforcing terminology consistency
• Style-guided translation — matching a target tone or register

The model ships with an 8,000-token context window, which is modest compared to general-purpose large language models but in line with other dedicated translation systems that process shorter, discrete text segments rather than long documents in a single pass.

Pricing and availability

Hy-MT2-30B-A3B is available through Tencent Cloud at $0.074 per 1 million input tokens and $0.295 per 1 million output tokens. OpenRouter lists the model as newly added, with uptime and latency data still being collected — the platform notes there is "not enough uptime data to display yet" and no throughput figures are currently available. The listing shows a release date of August 20, 2026.

No independent benchmark scores were provided in Tencent's release materials or the OpenRouter listing. Tencent has not disclosed the training data composition, training cutoff date, or comparative evaluation results against other translation systems such as Google Translate's NMT models or DeepL's proprietary engines.

What this means

Hy-MT2-30B-A3B is a specialized tool, not a general-purpose chatbot competitor. By using a mixture-of-experts design with only 3B active parameters out of 30B total, Tencent is betting that translation-specific efficiency — cheap inference on a narrow task — matters more than raw scale for this use case. The pricing, at roughly a third of a cent per 1,000 output tokens, positions it as a low-cost option for high-volume translation pipelines, particularly for enterprises already inside the Tencent Cloud ecosystem.

The explicit support for Chinese dialects and minority languages signals a domestic-market focus, distinguishing it from Western translation APIs that prioritize major world languages. The lack of published benchmarks makes it difficult to independently verify translation quality against incumbents like DeepL or Google's translation stack, so buyers evaluating this model for production use should run their own comparative tests before committing volume. The 8K context ceiling also means it is best suited for sentence- or paragraph-level translation rather than long-document workflows without chunking.