Skip to main content
Back

Tencent's Hy-MT2 Beats Commercial Translation APIs at 440MB on Device

Tencent open-sources Hy-MT2, a family of translation-specialized LLMs that beat DeepSeek and commercial APIs across 33 languages

  • Tencent open-sourced Hy-MT2, a family of 3 translation-specialized LLMs (1.8B, 7B, 30B-A3B MoE) supporting 33 languages.
  • The 7B and 30B models outperform DeepSeek-V4-Pro and Kimi K2.6 in fast-thinking mode; the 1.8B beats Microsoft Translator and Doubao overall.
  • 1.25-bit quantization via AngelSlim compresses the 1.8B to just 440MB with 1.5x inference speedup -- enabling true offline on-device translation.
  • Models follow complex translation instructions: terminology pinning, style control, delimiter preservation, structured data (JSON/XML) translation, and more.
  • Training uses a novel Family-Centric Post-training pipeline with language-family-specific RL (GRPO) and multi-teacher distillation across Western European, East Asian, and Middle Eastern language groups.
  • IFMTBench, a new open-source benchmark for translation instruction-following, is released alongside the models on GitHub.

Tencent's Hunyuan team just open-sourced Hy-MT2, a family of translation-specialized language models that punches well above its weight class. The lineup spans three sizes -- 1.8B, 7B, and a 30B Mixture-of-Experts model that activates only 3B parameters per inference -- and all three support 33 languages out of the box. Within days of launch, the 1.8B model hit #1 and the 30B-A3B hit #4 on Hugging Face's open-source trending leaderboard, with over 7,000 downloads in the first few days.

The "Fast-Thinking" Bet

The core design philosophy here is what Tencent calls "fast-thinking" translation. Standard LLMs used for translation tend to reason slowly -- they process full semantics before generating output. Traditional LLM translation uses a "slow thinking" approach, while Hy-MT2 introduces a "fast-thinking" paradigm that reacts like a professional human translator, cutting unnecessary reasoning overhead. The practical result: faster inference without sacrificing quality, which matters enormously for real-time applications like video subtitles, live meetings, or mobile apps.

The payoff is real. The 7B and 30B models outperform open-source models such as DeepSeek-V4-Pro and Kimi K2.6 in fast-thinking mode, while the lightweight 1.8B model also surpasses mainstream commercial APIs from providers such as Microsoft and Doubao overall. That last point is particularly striking -- a sub-2B model beating cloud translation services is not something you see every day.

How It Was Built

The training pipeline is a three-stage process described in the Hy-MT2 technical paper. Each stage builds on the previous one:

  1. MT-oriented Mid-training: Starting from a general Hunyuan pretraining model, the team continued training on approximately 1 trillion tokens of multilingual translation data, covering both monolingual corpora and parallel translation pairs across general, domain-specific, and real-world scenarios.
  2. Family-Centric Post-training (FCPT): Rather than mixing all languages together, the team split training into language-family branches (Western European, East Asian, Middle Eastern, etc.). Each branch got its own specialized teacher model -- called a "Chimera Teacher" -- built by fusing outputs from multiple Hunyuan reference models. Each branch was then fine-tuned with GRPO (a reinforcement learning algorithm), using a 5-dimensional quality reward system scoring terminology, accuracy, linguistic conventions, style, and instruction-following. Finally, a cross-family distillation step merged all the family-specific experts back into a single unified model.
  3. Quantization: Using Tencent's own AngelSlim toolkit, the team produced variants from FP16 all the way down to 1.25-bit. When paired with AngelSlim's 1.25-bit extreme quantization for on-device deployment, the lightweight 1.8B model requires only 440 MB of storage and achieves a 1.5x inference speedup.

The 1.25-bit quantization is worth dwelling on. It uses a technique called "Sherry" -- a sparse ternary scheme that packs four weights into 5 bits using a 3:4 sparsity pattern. Real-world business deployment exposed the limitations of the previous Hy-MT1.5 in on-device efficiency: the 4-bit quantized version still required more than 1GB of storage, and its inference speed was insufficient for low-latency scenarios. To address this, Hy-MT2 implements 1.25-bit extreme quantization based on Hunyuan's AngelSlim technology.

Benchmark comparison of Hy-MT2 models across WildMTBench, WMT25, Minority-Mandarin, FLORES-200, and instruction-following tasks

What It Actually Does Well

Multi-dimensional evaluations show that Hy-MT2 delivers outstanding performance across general, real-world business, domain-specific, and instruction-following translation tasks. But the most interesting capability is instruction-following during translation -- something most translation APIs completely ignore.

The model understands and executes complex translation constraints passed in natural language. Supported instruction types include:

  • Terminology pinning: Provide a glossary of terms and the model will use them consistently throughout the translation
  • Style control: Specify formal, casual, literary, or domain-specific register
  • Delimiter preservation: Special characters like ## or template variables like ${var} are kept intact -- critical for software localization
  • Structured data translation: Translate only user-visible text in JSON or XML, leaving keys and code untouched
  • Background context: Provide reference material so the model translates with full situational awareness
  • Personalization: Multi-step instruction lists for highly specific output requirements

On the team's own IFMTBench benchmark (which they also open-sourced), Hy-MT2-30B-A3B achieves the best overall IFMTBench score among small-to-medium-sized models, with consistent gains especially on complex multi-constraint instructions. The 30B model approaches GPT-5.5 and Gemini 3.1 Pro on translation instruction-following despite being a fraction of their size.

Where It Falls Short

The models are optimized specifically for translation. On general instruction-following benchmarks like Multi-IF (multi-turn conversations), Hy-MT2 trails comparable general-purpose models in later turns. The paper notes that "the main advantage of Hy-MT2 lies in translation-specific instruction following rather than general multi-turn instruction following." If you need a Swiss-army-knife LLM, this isn't it. Also, the 1.8B model's general instruction-following scores on IFEval (80.22) lag behind similarly-sized general models, a natural tradeoff for specialization.

Getting Started

All models are available on GitHub and Hugging Face. Hy-MT2 is released in multiple precision formats, including 1.25-bit, 2-bit, 4-bit, 8-bit, and FP16. The team supports four inference backends: Transformers, vLLM, SGLang, and llama.cpp. Here's the minimal Transformers usage for the 30B model:

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
model_path = "tencent/Hy-MT2-30B-A3B"
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_path, dtype=torch.bfloat16, device_map="auto", trust_remote_code=True
)
prompt = "Translate the following text into French. Only output the result:\n\nHello, how are you?"
messages = [{"role": "user", "content": prompt}]
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))

Hy-MT2 provides a complete model training pipeline, supporting both full-parameter fine-tuning and LoRA fine-tuning, as well as multiple DeepSpeed ZeRO configurations and LLaMA-Factory integration. The MoE architecture of the 30B-A3B model is worth noting: it has 30B total parameters but only 3B activate per inference, theoretically tapping 30B's knowledge capacity at 3B compute cost. That makes it surprisingly affordable to serve.

The Bigger Picture

The machine translation space has long been dominated by closed commercial APIs. What Hy-MT2 demonstrates is that a purpose-built, open-source translation model -- trained with a carefully designed RL reward system and language-family-aware distillation -- can match or beat those APIs at a fraction of the inference cost. The 440MB on-device variant in particular opens up offline translation on mobile hardware that would have been unthinkable a year ago.

In this release, Tencent also open-sources IFMTBench, a benchmark for evaluating translation instruction-following capabilities. That's a meaningful contribution on its own -- there's been no good standardized way to measure whether a model actually follows translation constraints, and this fills that gap. Tencent Hy is also officially partnering with WMT26 for the "Video Subtitle Translation Task," with participants using the Hy-MT model series having the chance to win special awards sponsored by Hunyuan. The community signal is clear: translation as a first-class ML problem, not an afterthought.