Idea
Multilingual translation models delivering faster, smaller, and more accurate real-world language solutions across 33 languages.
Research Paper
Core Innovation
This paper introduces Hy-MT2, a family of multilingual translation models optimized for speed, efficiency, and instruction-following across 33 languages. It advances prior work by combining model scaling with extreme quantization for on-device deployment and demonstrating superior performance over leading open-source and commercial models in diverse real-world scenarios.
Why It Matters
Global businesses and developers require fast, accurate, and scalable multilingual translation to operate efficiently across diverse markets. Hy-MT2 reduces storage and latency for on-device use while outperforming existing open-source and commercial APIs, enabling broader adoption in real-world and domain-specific applications. This scalability and efficiency transform workflows by making high-quality translation accessible on edge devices and cloud platforms.
Market Size (TAM)
$10–20B TAM for multilingual translation platforms; $2–5B SAM from global enterprises and mobile developers. Driven by globalization and mobile edge computing adoption.
Potential Customers & Pain Points
- Global enterprises – Need scalable accurate multilingual translation
- Mobile app developers – Require lightweight fast on-device models
- Localization services – Demand domain-specific and instruction-following translation
- Cloud providers – Seek cost-effective high-performance translation APIs
Business Model
Subscription-based API access for cloud and enterprise customers; licensing for on-device deployment; custom domain adaptation services.
Competitive Landscape
- DeepSeek-V4-Pro
- Kimi K2.6
- Microsoft Translator
- Google Translate API
- Amazon Translate
Implementation Challenges
- Integration complexity with existing localization pipelines
- Maintaining translation quality across diverse domains and languages
- Competition from established commercial APIs with large user bases
Validation Strategy
- Benchmark against leading open-source and commercial translation APIs in real-world business scenarios
- Pilot deployments with mobile app developers for on-device inference speed and storage benefits
- Customer feedback loops for domain-specific translation quality improvements
Research Paper Overview
Hy-MT2: A Family of Fast, Efficient and Powerful Multilingual Translation Models in the Wild
Summary
Hy-MT2 offers three scalable multilingual translation models (1.8B, 7B, 30B-A3B MoE) supporting 33 languages with strong instruction-following capabilities. The 1.8B model enables on-device deployment with extreme quantization, reducing storage to 440 MB and improving inference speed by 1.5x. Evaluations show superior performance over leading open-source and commercial translation solutions across diverse real-world and domain-specific tasks.