Hy-MT2-30B-A3B

Hy-MT2-30B-A3B sounds more intimidating than it really is. The name is full of model jargon: 30B, A3B, MoE. But the part that caught my attention had nothing to do with parameter counts. It was the price.
Advertisement 728 × 90
CompanyTencent
Context8K
Released2026-08
Updated2026-09-02

Hy-MT2-30B-A3B Overview

Hy-MT2 is Tencent Hunyuan’s multilingual translation model family, released as open source in May 2026. The lineup includes 1.8B, 7B, and 30B-A3B versions.

The 30B-A3B model uses a Mixture-of-Experts, or MoE, architecture.

It has roughly 30 billion parameters in total, but only about 3 billion are active during each translation task. That gives the model more capacity without forcing every request to run through the full 30 billion parameters.

It supports translation across 33 languages, including Chinese-English, English-to-other-language pairs, and several minority languages.

You can also give it fairly specific instructions: lock certain terms, control the writing style, preserve formatting, or limit sentence length.

Compared with Hy-MT1.5, the biggest change is not that it suddenly “knows more.” It follows instructions more reliably. For anyone trying to put a translation model into a real workflow, that matters more than another flashy feature.

Hy-MT2-30B-A3B Pricing

PlanPriceDescription
Input $0.074 Charged by input token usage
Output $0.295 Charged by output token usage

Hy-MT2-30B-A3B Key Features

1. Domain-Specific Translation

The 30B-A3B model has been optimized for eight professional domains, including finance, law, politics, and education.

On DomainMTBench, it scored 97.08 in finance and 89.15 in law using XCOMET.

Professional translation usually breaks in small ways before it breaks in obvious ones.

A technical term is translated one way on page three and a different way on page nine. A legal sentence is technically translated word for word, but the logical structure gets muddled. A financial term is correct in isolation but inconsistent across the full document.

Those are the kinds of problems that matter in real work.

That is also where a larger, more specialized model starts to earn its keep.

2. Better Instruction Following

This is one of the clearest improvements over Hy-MT1.5.

You can give the model a rule like:

“Keep the translation concise and limit each sentence to 15 words.”

You can also provide a glossary and force certain terms to use a fixed translation.

Real projects usually come with several rules at once.

Keep the formatting. Preserve punctuation. Do not translate field names. Keep product names unchanged. Use the same tone throughout.

None of those requirements is difficult by itself. The problem starts when they are all combined.

Hy-MT2 was built to handle that kind of instruction stack more reliably. Tencent also released IFMTBench to measure this capability, and the 30B-A3B model scored 84.69 on the complex-instruction test.

For production translation, I find this kind of result more useful than another generic “fluency” score. It tells you whether the model can actually stay inside the rules you give it.

3. Performance on Real-World Translation Tasks

Hy-MT2-30B-A3B scored 89.87 on WildMTBench using GEMBA, ahead of Gemini 3.1 Pro in that benchmark.

WildMTBench is designed to look more like real translation work, where the input is less clean and predictable than standard test data.

That matters.

Actual business documents rarely look like benchmark examples. Technical files mix abbreviations, table fields, half-translated phrases, and inconsistent formatting. Commercial documents contain industry jargon. Contracts are full of nested clauses and long sentences.

A model that performs well on polished test sentences can still fall apart on that kind of material.

The WildMTBench result suggests that 30B-A3B is not just good at controlled benchmark conditions.

4. What Changed From Hy-MT1.5

The most noticeable upgrade is instruction following.

Hy-MT1.5 was more likely to miss a rule when several translation requirements were combined. Hy-MT2 is more consistent with terminology, formatting, length limits, and style constraints.

Minority-language translation also improved.

In tests covering languages such as Tibetan, Uyghur, and Mongolian, the 30B-A3B model reached an XCOMET score of 62.44, higher than Gemini 3.1 Pro in the same comparison.

The Hy-MT2 family also pushes quantization quite aggressively.

That matters less for the 30B-A3B model itself, which is better suited to server or cloud deployment. But the same work carries down to the smaller models. The 1.8B version can be compressed to around 440MB and run locally on a phone.

The product strategy is easy to see: the larger model focuses on professional translation quality, while the smaller one is pushed toward on-device use.

Summary

What makes Hy-MT2-30B-A3B interesting is not the “30B” label by itself, and it is not one benchmark score.

It is the combination.

The model performs well on finance, legal, and other professional material. It follows complex translation instructions more reliably than the previous generation. Its MoE setup keeps the active parameter count at around 3 billion per request.

Then Tencent prices the API exactly the same as the 7B model.

That is the part I did not expect.

For companies translating contracts, financial documents, technical material, or cross-border content every day, I would look at the 30B-A3B version first. Teams working with Tibetan, Uyghur, Mongolian, and similar language pairs also have a good reason to test it.

The limitations are clear too.

An 8,192-token context window is not especially generous for a model aimed at professional document translation. Long contracts, research papers, and large technical manuals will still need to be split into chunks.

Function Calling is not supported either.

So this is not an AI agent platform or a general-purpose assistant. It is a translation model, and the boundaries are pretty strict.

Recommended for

localization teams, cross-border businesses, and companies that regularly work with legal, financial, or technical documents. It also makes sense for organizations handling minority-language translation at scale.

Not recommended for

 anyone who only needs to translate the occasional email, sentence, or casual conversation. Free online translation tools are more than enough for that, and even the 7B model would be plenty.

The 30B-A3B starts to make sense when translation stops being an occasional task and becomes part of the actual business workflow.

Comments (0)

Leave a comment

Advertisement 728 × 90

Tencent Model Comparison

Model Context Pricing API Released Global Heat
Hy-MT2-30B-A3B
8K YES 2026-08
8K YES 2026-08
1M YES 2026-08
262K YES 2026-07
59/100

Similar Models

Tencent: Hy3
59
Tencent Hunyuan’s advanced MoE language model optimized for reasoning, coding, long-context understanding, and AI Agent applications.
Tencent
Hy-MT2-7B
Hy-MT2-7B is easy to understand because it does not try to be everything at once. It is not built for chatting, web search, or code generation. Its job is translation. That narrow focus works in its favor, especially when you need consistent terminology, longer translations, or strict formatting.
Tencent
Hy4 preview
On August 28, Tencent’s Hunyuan team released its next-generation large language model, Hy4 preview, and open-sourced the model weights at the same time. Tencent’s pitch is pretty clear: “Built for productivity.”
Tencent
Claude Opus 5
100
Anthropic’s flagship Claude model built for complex reasoning, AI Agents, software development, and enterprise knowledge workflows.
Anthropic
GPT-5.6 Luna
99
An efficient GPT-5.6 series model optimized for speed, cost efficiency, and intelligent performance across large-scale AI applications.
OpenAI
GPT-5.5 Pro
99
I’ve been using GPT-5.5 Pro with a few colleagues for the past several weeks, mostly on the kinds of jobs where regular chatbots tend to fall apart: long documents, messy research questions, code debugging, and tasks that need more than one or two steps of reasoning.
OpenAI
GPT-5.5
99
OpenAI flagship model, a multimodal AI natively supporting text, images, and audio.
OpenAI
Claude Opus 4.8
98
Claude Opus 4.8 is Anthropic's most powerful Opus-series model, featuring multimodal input, reasoning, and a 1M-token context window, excelling in complex reasoning and coding.
Anthropic

Related Tools

Related News