Hy-MT2 – Tencent Hunyuan's Open-Source Translation Model Family

1. What Is Hy-MT2

Hy-MT2 is Tencent Hunyuan's open-source next-generation machine translation model family in 1.8B, 7B, and 30B-A3B sizes—33-language mutual translation plus five ethnic/minority and dialect pairs. The 7B and 30B-A3B variants lead open models on FLORES-200 and real-world evals, beating some closed commercial models many times larger; the 1.8B lightweight tier uses proprietary 1.25-bit quantization down to ~440MB for on-phone offline inference. Hy-MT2 advances both quality and edge deployment for open translation.

Image source: Official article

Technical positioning and domain: Large-scale multilingual MT with decoder-only LLM architecture—instruction-tuned and multi-task trained—covering cloud high-performance and mobile offline extremes rarely combined in OSS MT.

Development background: From Tencent Hunyuan with long MT heritage. Goals: close the OSS quality gap in production and satisfy rigid mobile offline demand via architecture, quantization, and training innovations.

Core value: Resolves the performance–efficiency tradeoff—7B/30B-A3B rival or beat closed APIs cost-effectively; 1.8B quantized runs locally without network. Strong instruction following suits finance, legal, medical verticals.

Technical characteristics: Multi-size lineup; AngelSlim 1.25-bit quantization; multi-scenario joint training; translation instruction tuning; 30B-A3B MoE (~3B active) for cheaper inference; 1.8B on Apple/Qualcomm/MediaTek chips ~1.5× faster than prior gen.

2. Key Features

• Multilingual translation: 33 languages plus Tibetan, Uyghur, Mongolian, Korean (ethnic script context), Yi, etc.—strong on low-resource pairs vs other OSS models.

• Voice input translation: WeChat mini program "Tencent Hy Translate" combines ASR + MT for spoken cross-language input—travel and meetings.

• Custom translation style: Terminology, format, brevity via instructions—financial terms, academic structure preservation; strong IFMTBench instruction following.

• On-device offline: 1.8B AngelSlim ~440MB download—local inference offline on phones with second-scale latency on mainstream SoCs.

• Word analysis and comparison: Tap words for definitions, examples, synonyms; switch styles to compare phrasing.

• Vertical domain tuning: Finance, legal, medical, education, and four more—GEMBA scores near major commercial APIs without extra training for many docs.

3. How to Use

1. Requirements: Mini program needs no special hardware. Self-host: 1.8B ~4GB VRAM (quantized on phone); 7B 16GB+; 30B-A3B 40GB+. Linux/macOS, Python 3.8+, PyTorch 2.0+.

2. Online: WeChat search "腾讯Hy Translate"—pick languages, type or speak; personalize style in settings.

3. Offline: In app settings download 1.8B quantized (~440MB); switch offline mode for no-network use.

4. OSS deploy: Weights on GitHub/Hugging Face—pick size; load via transformers; 1.8B use bitsandbytes; 7B/30B-A3B prefer vLLM or TensorRT-LLM.

5. Config tips: do_sample=False for deterministic MT; add domain system_prompt for vertical docs; chunk long text (~512 tokens) for coherence.

6. Best practices: Download models on stable Wi‑Fi; specify terminology; manual source language if auto-detect fails.

4. Pros and Cons

5. Comparison with Similar Tools

Selection advice: Mobile offline—Hy-MT2 uniquely viable. Enterprise OSS MT with instructions—7B/30B-A3B best ROI. Extreme 200-language coverage—NLLB-200. Real-time speech cloud—GPT-Realtime-Translate but closed/costly.

6. Editor's Review

Hy-MT2 innovates on three fronts: systematic multi-size design phone-to-cloud; AngelSlim 1.25-bit enabling credible offline phone MT; MoE 30B-A3B for affordable high quality.

7B/30B-A3B beating much larger closed models on FLORES-200 and vertical GEMBA lowers enterprise MT cost. Offline 1.8B solves travel/meeting connectivity gaps.

Audience: mobile devs, enterprise MT teams, MT researchers, everyday users via mini program.

Open strategy plus Tencent backing should expand languages, domains, voice, and long-context support.

— Best-in-class open MT on performance, edge deploy, features, and license friendliness; minor doc/domain gaps don't outweigh strengths.

7. Use Cases

• Daily cross-language chat: 33 languages + voice in mini program; offline mode for unstable networks.

• Professional documents: Finance/legal/medical/education with terminology and format control.

• Offline travel and field work: 440MB local model on planes, remote areas, subway.

• Creator localization: Custom tone for translating references into platform voice.

• Business communication: Meetings, email, trade—word glossaries for jargon-heavy dialogue.

8. FAQ

Q: Does 1.8B offline cover all 33 languages?
A: All major pairs supported; low-resource/dialect quality below cloud—prefer cloud for rare pairs when online.

Q: 30B-A3B architecture and cost?
A: MoE 30B total, ~3B active per forward—~3–5× faster than dense 30B class; ~40GB VRAM (single A100/H100).

Q: Integrate into my app / API?
A: Self-host weights with transformers; wrap your API; accelerate with vLLM/TensorRT-LLM—no official Tencent cloud MT API for Hy-MT2.

Q: Vertical quality without fine-tuning?
A: Pre-tuned on eight domains near commercial APIs; add glossary/style prompts for niches.

Q: How does 1.25-bit quantization work?
A: AngelSlim mixed precision + distillation; FLORES BLEU drop ≤2 points; subjective gap minimal vs full precision.

Q: Non-Chinese–English pairs?
A: Yes—JA/KO/FR/DE etc. mutual; pick source/target explicitly.

Q: Phone inference speed?
A: ~0.5–1s on Apple A15; ~0.8–1.5s on Snapdragon 8 Gen 2—~1.5× vs Hy-MT1; download model locally first.

9. Project Links

• GitHub: https://github.com/Tencent-Hunyuan/Hy-MT2
• Hugging Face: https://huggingface.co/collections/tencent/hy-mt2
• Tencent AI Studio: https://aistudio.tencent.com/llm/en?tabIndex=0