Qwen-MT: AI-Powered Machine Translation

Qwen-MT (Machine Translation) is Alibaba's dedicated translation model — purpose-built for high-quality multilingual translation across 92 languages covering 95%+ of the world's population. Released in July 2025, it's trained on the Qwen3 architecture with reinforcement learning specifically optimized for translation accuracy and fluency. Available in three tiers (Flash, Plus, Lite) with API pricing starting at $0.50 per million tokens, it outperforms GPT-4.1-mini and Gemini-2.5-Flash on translation benchmarks while competing with GPT-4.1 at a fraction of the cost.

Navigate this guide:

• What Is Qwen-MT?
• 92 Languages Supported
• Model Tiers: Flash, Plus, Lite
• Key Features
• Benchmarks vs GPT-4 and Gemini
• API Pricing
• API Quick Start
• Use Cases
• vs Google Translate vs DeepL
• Limitations
• FAQ

What Is Qwen-MT?

While general-purpose LLMs like Qwen 3.5 can translate text, they're not optimized for it — they can produce inconsistent terminology, miss nuances, or hallucinate content not in the source. Qwen-MT solves this by being a translation-specialized model: same Qwen3 architecture, but fine-tuned with reinforcement learning specifically for translation quality, consistency, and domain adaptation.

The result is a model that produces translations with fewer errors, more consistent terminology, and better handling of domain-specific jargon than general-purpose alternatives — while being significantly cheaper to run.

92 Languages Supported

Qwen-MT supports 92 languages covering 95%+ of the global population. Major language families include:

• European — English, Spanish, French, German, Italian, Portuguese, Dutch, Polish, Czech, Romanian, Swedish, Danish, Norwegian, Finnish, Greek, Hungarian, Bulgarian, Croatian, Slovak, Lithuanian, Latvian, Estonian, Slovenian, Maltese, Irish, Catalan, Basque, Galician
• East Asian — Chinese (Simplified & Traditional), Japanese, Korean
• South/Southeast Asian — Hindi, Bengali, Tamil, Telugu, Thai, Vietnamese, Indonesian, Malay, Tagalog, Burmese, Khmer, Lao
• Middle Eastern / African — Arabic, Turkish, Persian, Hebrew, Urdu, Swahili, Amharic
• Slavic — Russian, Ukrainian, Serbian, Bosnian, Macedonian, Belarusian

Model Tiers

Key Features

Terminology Intervention

Define custom glossaries and the model will respect them throughout the translation. Critical for brands, technical documentation, and regulated industries where specific terms must be translated consistently.

Domain Prompts

Specify the domain context ("medical", "legal", "gaming", "e-commerce") and the model adapts its vocabulary, register, and style accordingly. No fine-tuning needed — just a prompt parameter.

Translation Memory

Feed previously approved translations as context, and the model will maintain consistency with your existing translated corpus. Essential for large-scale localization projects.

Reinforcement Learning Optimized

Unlike general LLMs that translate as a side effect of language understanding, Qwen-MT was explicitly optimized through RL for translation metrics: faithfulness, fluency, and terminological consistency.

Benchmarks

Qwen-MT excels particularly on low-resource language pairs and domain-specific content where general models tend to struggle.

API Pricing

For comparison, GPT-4.1 costs $2.00/$8.00 per 1M tokens. At the Lite tier, Qwen-MT is 4-16x cheaper than comparable GPT-4 translation.

API Quick Start

Qwen-MT uses the DashScope API with an OpenAI-compatible format:

import openai

client = openai.OpenAI(
    base_url="https://dashscope-intl.aliyuncs.com/compatible-mode/v1",
    api_key="your-dashscope-api-key"
)

response = client.chat.completions.create(
    model="qwen-mt-plus",
    messages=[
        {"role": "system", "content": "You are a professional translator. Domain: legal."},
        {"role": "user", "content": "Translate to Spanish: The party of the first part hereby agrees..."}
    ]
)
print(response.choices[0].message.content)

With Terminology Intervention

# Add a glossary parameter
response = client.chat.completions.create(
    model="qwen-mt-plus",
    messages=[
        {"role": "system", "content": "Translate EN→DE. Glossary: 'cloud computing'='Cloud-Computing', 'machine learning'='maschinelles Lernen'"},
        {"role": "user", "content": "Our cloud computing platform uses machine learning for optimization."}
    ]
)

Use Cases

• E-commerce localization — translate product listings, reviews, and support content across markets at scale with the Lite tier
• Legal/medical translation — use Plus tier with terminology intervention for regulated content requiring exact terminology
• Software localization — translate UI strings, documentation, and release notes with translation memory for consistency
• Content marketing — adapt blog posts, landing pages, and campaigns for international audiences
• Customer support — real-time chat translation for multilingual support teams
• Academic translation — translate research papers and abstracts maintaining technical accuracy

Qwen-MT vs Traditional Translation

The key advantage of Qwen-MT over traditional NMT (Neural Machine Translation) services is full document-level context: it reads the entire document before translating, maintaining consistency of pronouns, references, and terminology throughout — something sentence-level translation APIs struggle with.

Limitations

• API only — no open-weight model available for local deployment
• 92 languages — fewer than Google Translate's 133, missing some very low-resource languages
• Latency — LLM-based translation is slower than traditional NMT for simple sentences
• No real-time speech translation — for speech translation, combine with Qwen3-ASR and Qwen3-TTS, or use Qwen3-Omni

FAQ

Is Qwen-MT a separate product from Qwen Chat?

Qwen-MT is available both as a standalone API (via DashScope) and as a feature within Qwen Chat. The API gives you programmatic access with glossaries and domain prompts; Qwen Chat offers a simpler interface for one-off translations.

Can I use it for real-time chat translation?

Yes. The Lite and Flash tiers are fast enough for real-time chat scenarios. For voice-to-voice translation, pair it with Qwen3-ASR for input and Qwen3-TTS for output.

How does it compare to just asking Qwen 3.5 to translate?

Qwen 3.5 can translate, but Qwen-MT is specifically optimized for translation: better consistency, lower hallucination rate, glossary support, and 2-5x cheaper per token for translation workloads.

Is the model open-source?

No. Qwen-MT is currently available as an API service only. The underlying Qwen3 models are open-source, but the translation-specific fine-tuning is proprietary.