# Hy-MT2-30B-A3B ![Organization](https://img.shields.io/badge/Org-Hungyuan-6366f1) ![Params](https://img.shields.io/badge/Params-30B%20MoE%20(3B%20active)-8957e5) ![Context](https://img.shields.io/badge/Context-262K%20tokens-1f6feb) ![License](https://img.shields.io/badge/License-Apache%202.0-2da44e) ![Modality](https://img.shields.io/badge/Modality-Text-8957e5) ![Released](https://img.shields.io/badge/Released-May%202026-555) > **Quick answer:** Hy-MT2-30B-A3B is the flagship member of Tencent Hunyuan's Hy-MT2 family — a 30B-parameter Mixture-of-Experts model (3B active parameters, 128 experts with 8 active + 1 shared) supporting 33 languages, released under Apache 2.0. It leads its family on FLORES-200's full translation matrix at 87.47 (XCOMET-XXL), while running inference at roughly the cost of a 3B dense model thanks to its MoE routing. ## At a Glance **Where Hy-MT2-30B-A3B leads** - Best-in-family score on FLORES-200 (87.47 XCOMET-XXL) and on domain-specific/real-world translation evaluations (Tencent's in-house DomainMTBench and WildMTBench) - MoE efficiency: only 3B parameters active per forward pass (128 experts, 8 active + 1 shared) despite 30B total — inference cost closer to a small dense model than its parameter count suggests - Outperforms Tower-Plus-72B (a much larger dedicated open translation model) across every reported general-translation benchmark **Where it lags** - Narrowly behind the smaller dense Hy-MT2-7B on WMT25 specifically (62.89 vs 63.86 XCOMET-XXL) — the family's one benchmark where bigger doesn't win - Behind frontier general-purpose models like Gemini 3.1 Pro on raw translation quality across the board - General instruction-following (IFBench, IFEval, MaXIFE, Multi-IF) is weaker than translation-specific instruction following — it's tuned for translation, not general-purpose chat **Best for:** Deployments wanting the highest overall translation quality in the Hy-MT2 family while keeping inference cost close to a small dense model, via MoE's sparse activation. ## What Hy-MT2-30B-A3B Is Hy-MT2-30B-A3B is the flagship size (of 1.8B, 7B, 30B-A3B) in Tencent Hunyuan's second-generation "fast-thinking" multilingual translation model family, described in the technical report "Hy-MT2: A Family of Fast, Efficient and Powerful Multilingual Translation Models in the Wild." Unlike its two dense siblings, it uses a Mixture-of-Experts architecture (128 experts, 8 active plus 1 shared expert per token, ~3B active parameters), trained via the same three-stage Family-Centric Post-training (FCPT) pipeline: MT-oriented mid-training on ~1T tokens of translation-related data, followed by Reference-Guided On-Policy Distillation and Family-specific RL training organized around language families (Western European, East Asian, Middle Eastern right-to-left), then a final Cross-family distillation step into a unified model. Like its siblings, Hy-MT2-30B-A3B is evaluated across four dimensions: general translation (FLORES-200, WMT25, Mandarin⇔Minority), real-world business scenarios (Tencent's own WildMTBench), domain-specific translation (Tencent's own DomainMTBench, covering finance/law/medicine/technology/politics/education), and translation instruction-following (IFMTBench, plus general instruction benchmarks IFBench/IFEval/MaXIFE/Multi-IF). It's the strongest all-around model in the family, leading on FLORES-200 and the in-house domain/real-world suites, while its MoE design keeps per-token compute close to a 3B dense model. ## Specifications | Field | Value | |---|---| | Organization | Hungyuan (Tencent) | | Architecture | Mixture-of-Experts (128 experts, 8 active + 1 shared) | | Total parameters | 30B | | Active parameters | ~3B | | Context length | 262,144 tokens | | Supported languages | 33 | | License | Apache 2.0 | | Release date | May 2026 | | Modality | Text | | Quantized variants | FP8, GGUF (multiple quant levels) | ## Pricing Open weights under Apache 2.0 — self-hosted via transformers, vLLM, SGLang, or llama.cpp (GGUF builds available). No API list price; cost is your own inference infrastructure. ## Public Benchmark Scores | Benchmark | Score | Source | Date | |---|---|---|---| | [FLORES-200](/benchmarks/meta-ai/flores-200) | 87.47 (XCOMET-XXL) | Hy-MT2 technical report | 2026-05 | | [WMT25](/benchmarks/wmt/wmt25) | 62.89 (XCOMET-XXL) | Hy-MT2 technical report | 2026-05 | *Hy-MT2-30B-A3B also reports scores on Tencent's own in-house evaluation suites (WildMTBench, DomainMTBench, IFMTBench, Mandarin⇔Minority testset) and on general instruction-following benchmarks (IFBench, IFEval, MaXIFE, Multi-IF) — not added as Benchgen leaderboard entries since these either lack an independent public benchmark page (WildMTBench/DomainMTBench/IFMTBench are Tencent-constructed, non-independently-reproducible test sets) or measure general instruction-following rather than translation capability specifically.* ## Hy-MT2-30B-A3B vs Alternatives | Model | FLORES-200 (XCOMET-XXL) | WMT25 (XCOMET-XXL) | License | Params | |---|---|---|---|---| | [Hy-MT2-1.8B](/models/hungyuan/hy-mt2-1-8b) | 79.77 | 50.30 | Apache 2.0 | 1.8B | | [Hy-MT2-7B](/models/hungyuan/hy-mt2-7b) | 86.89 | 63.86 | Apache 2.0 | 7B | | **Hy-MT2-30B-A3B** | **87.47** | 62.89 | Apache 2.0 | 30B (3B active) | Hy-MT2-30B-A3B is the strongest all-around model in the family, only narrowly ceding the WMT25 lead to the smaller dense Hy-MT2-7B. ## Related Models & Benchmarks - Compare: [Hy-MT2-1.8B](/models/hungyuan/hy-mt2-1-8b) · [Hy-MT2-7B](/models/hungyuan/hy-mt2-7b) - Benchmarks: [FLORES-200](/benchmarks/meta-ai/flores-200) · [WMT25](/benchmarks/wmt/wmt25) ## Frequently Asked Questions

What is Hy-MT2-30B-A3B?

Hy-MT2-30B-A3B is Tencent Hunyuan's flagship translation-specialist model, a 30B-parameter Mixture-of-Experts model (3B active parameters) supporting 33 languages, released under Apache 2.0.

Is Hy-MT2-30B-A3B open source?

Yes — it's released under the Apache 2.0 license with weights on Hugging Face and ModelScope, including FP8 and GGUF quantized variants.

What does the "A3B" in Hy-MT2-30B-A3B mean?

"A3B" denotes the model's active-parameter count under its Mixture-of-Experts architecture — 30B total parameters, but only about 3B active per forward pass (128 experts, 8 active plus 1 shared).

--- *Specs from Tencent Hunyuan's official Hy-MT2-30B-A3B Hugging Face model card and technical report (arXiv:2605.22064). Last updated 2026-08-31.*

Hy-MT2-30B-A3B — Training jobs

Training jobs