Hy-MT2: A Family of Fast, Efficient and Powerful Multilingual Translation Models in the Wild
Mao Zheng, Zheng Li, Tao Chen, Bo Lv, Mingrui Sun, Mingyang Song, Jinlong Song, Hong Huang, Decheng Wu, Hai Wang, Yifan Song, Yanfeng Chen, Guanwei Zhang
cs.CL
2026-05-21
Hy-MT2 covers 33 languages at 1.8B, 7B and 30B-A3B. The 7B hits 86.89 XCOMET on FLORES, 98% of Gemini 3.1 Pro thinking; 1.8B packs to about 440MB at 1.25-bit.
After Hy-MT1.5 shipped, the complaints clustered. Finance, law, and medical terms drifted. Web pages, meetings, and social posts arrived in messy formats. Users added constraints like keep this word, follow this template, and the model often ignored them. The 7B still lagged Gemini and GPT. On device, 4-bit still needed more than 1GB and was not fast enough. Hy-MT2 is aimed at those gaps: domains, in-the-wild text, translation instructions, an MoE 30B-A3B scale-up, and a 1.8B small enough for a phone.
It covers 33 languages at 1.8B, 7B, and 30B-A3B (about 3B active). The product is fast-thinking translation, not a slow general chat model.
Training starts from a Hunyuan pretrained checkpoint with about 1T tokens of MT-oriented mid-training: monolingual plus parallel data, covering general, domain, wild, and instruction cases.
Post-training does not mix every language family in one pot. Family-Centric Post-training splits branches. Each branch first runs Reference-Guided On Policy Distillation: several Hunyuan reference models plus dataset labels form a Chimera Teacher, the student samples online, and forward KL matches the fused preference. Then GRPO on that family. The reward is not a single holistic score. A rule filter zeros severe repetition and language mixing. Survivors go to an LLM judge that deducts under MQM: terminology, accuracy, linguistic conventions, style, instruction following. Start at 100, fatal errors go to 0, then a length penalty so short translations cannot farm the score.
A last Cross-family On Policy Distillation pours the family teachers plus a general instruction teacher into one student, now with reverse KL and extra general instruction data so the model does not forget how to follow non-MT prompts.
Precision options run from FP16 and 8/4-bit PTQ down to 2-bit SEQ QAT and 1.25-bit Sherry (3:4 sparse ternary) from AngelSlim. Ultra-low-bit QAT mixes LM loss with forward and reverse KL, weighted by the teacher's per-token confidence.
FLORES-200 all-pairs XCOMET: 79.77 / 86.89 / 87.47 for 1.8B / 7B / 30B-A3B, which the paper reads as 89.9%, 97.9%, and 98.6% of Gemini 3.1 Pro thinking (88.74). The 7B and 30B beat DeepSeek-V4-Pro non-thinking (83.29) and Kimi K2.6 non-thinking (84.05). Versus Hy-MT1.5-7B at 80.98, the 7B jumps nearly 6 points.
On WMT25 the 7B posts 63.86 / 71.21 / 82.24, up from 1.5's 61.59 / 68.85 / 75.91. The 30B-A3B GEMBA of 84.34 sits above Gemini thinking at 82.23. Mandarin-minority XCOMET is 62.05 for the 7B, a hair above Gemini's 61.11.
| Model | FLORES XX XCOMET | WMT25 XCOMET | Reference |
| Hy-MT2-1.8B | 79.77 | 50.30 | MT1.5-1.8B 78.40 / 53.08 |
| Hy-MT2-7B | 86.89 | 63.86 | MT1.5-7B 80.98 / 61.59 |
| Hy-MT2-30B-A3B | 87.47 | 62.89 | Gemini 3.1 ProT 88.74 / 57.58 |
DomainMTBench average for 30B-A3B is 95.04 / 93.73, best XCOMET in the table. WildMTBench: 7B 90.28 / 88.93; 30B GEMBA 89.25 versus Gemini 88.96. IFMTBench totals 83.14 (7B) and 84.69 (30B). General instruction following is less impressive: later Multi-IF turns lag similar-size open models.
The 1.8B beats Microsoft and Doubao on Chinese/English FLORES directions, WMT25, and wild text. All-pairs FLORES XCOMET is 79.77, a shade under Microsoft's 80.13. WMT25 XCOMET also falls from 1.5's 53.08 to 50.30, so the small model is not a clean upgrade. In the quant table, 1.8B-2bit FLORES all-pairs drops from 79.21 (BF16) to 76.31, and IFMTBench from 69.36 to 58.99. The abstract's 1.25-bit, 440MB, 1.5x versus 4-bit on Apple A15 never appears as quality numbers in the main quant table.
Specialist MT still earns a seat because of size, latency, and obeying user rules. A 7B fast-thinking model sitting next to Gemini thinking on FLORES is a deployment-cost argument. The 1.8B versus commercial APIs matters for on-device and air-gapped use, if you accept the WMT25 regression versus 1.5 and the instruction drop after 2-bit.
Family-wise teachers then a merge distillation is a training recipe against cross-family interference. It helps. It also costs more to run.
No standalone Limitations section. DomainMTBench, WildMTBench, and IFMTBench are in-house; annotation and splits are thin, and GEMBA is an LLM judge that can favor fluent outputs from a related model family. Many baselines are thinking-mode scores against a fast-thinking Hy-MT2. That comparison sells, it is not the same protocol. The 440MB 1.25-bit story lives in the abstract; the measured table stops at 2-bit. The 30B is MoE with 3B active, so compare it to other A3B models, not dense 30B. Multi-turn general instructions remain a weak spot if anyone tries to use it as a chatbot.