Tencent Open-Sources Hy-MT2: A 440MB On-Device AI Translation Model Running Fully Offline
Tencent's Hunyuan team open-sourced Hy-MT2, an ultra-compact 440MB on-device neural translation model running offline on mobile devices with 33 language support
On October 8, 2026, Tencent's Hunyuan AI team officially open-sourced Hy-MT2, an ultra-compact on-device neural machine translation model designed to run entirely offline on mobile and edge devices.
Image source: Simplifying AI (@simplifyinAI) / Tencent
Weighing in at just 440MB for its single-file model weights, Hy-MT2 can be deployed directly onto smartphones and local hardware, enabling fully self-contained multilingual translation without relying on external cloud infrastructure or active network connections.
440MB Footprint and Standalone Offline Translation
The defining technical advantage of Hy-MT2 is its compact footprint, allowing it to function seamlessly in zero-connectivity environments.
While mainstream large language model (LLM) translation workflows typically require gigabytes of memory and continuous remote API calls, Hy-MT2 operates comfortably within standard smartphone storage and RAM constraints at 440MB.
- Full Airplane Mode Capability: Translations process instantly in air travel, remote field locations, roaming-restricted areas, or air-gapped local networks without any internet connection.
- On-Device Data Privacy: Input text never leaves the user's local device, eliminating remote data transmission risks for confidential business documents and personal communications.
- Zero Cloud Subscription and Server Overhead: Developers and end users can deploy and integrate translation capabilities directly into standalone apps without recurring API costs or server maintenance.
33 Supported Languages and Benchmark Performance
Hy-MT2 provides translation capabilities across 33 languages, including Spanish.
According to internal benchmark evaluations released by the Tencent Hunyuan team, Hy-MT2 outperformed Microsoft's commercial translation API in their comparative test suite, delivering competitive translation accuracy despite its lightweight on-device architecture.
As with all proprietary vendor benchmarks, real-world translation quality and fluency may vary across specialized technical domains, informal colloquialisms, and lower-resource language pairs.
Open-Source Release and Edge Integration Potential
Tencent has made the model weights and documentation available via its official GitHub repository (Tencent-Hunyuan/Hy-MT2).
Developers can download the model weights to test locally or embed them directly into mobile messengers, offline travel translators, dictionary utilities, and localized productivity suites. The release highlights the expanding feasibility of self-hosted, privacy-preserving AI translation running natively on personal consumer hardware.
Sources
- Tencent Hunyuan Official GitHub: Tencent-Hunyuan/Hy-MT2
- Simplifying AI (@simplifyinAI): 2026-10-08 Hy-MT2 Release Announcement