Tencent Unveils Hy-MT2 Translation LLM with 7B, 30B-A3B, and 1.8B Models

Tencent released Hy-MT2, a free translation large language model on May 21. The company announced the release of the software update for artificial intelligence applications. Tencent made the model weights available to the public through open-source distribution channels.

Tencent releases open-source translation model supporting 33 languages globally

The Hy-MT2 suite includes three distinct models with different parameter counts. Users can access the 7B and 1.8B versions alongside a new 30B-A3B Mixture of Experts architecture. The smallest 1.8B variant utilizes AngelSlim 1.25bit extreme quantization technology. This compression technique reduces the model footprint to just 440MB.

Tencent claims superior performance across various translation tasks compared to existing open-source alternatives. The company states that the 1.8B model outperforms mainstream commercial translation APIs like Microsoft's service. Internal benchmarks also indicate the quantized version runs 1.5 times faster than the previous Hy-MT1.5 4-bit release on Apple A15 hardware.

Developers can download and use all three model variants globally without cost. The open-source weights are currently accessible through Hugging Face for immediate integration into translation workflows.

Performance metrics rely entirely on internal testing conducted by Tencent. Independent verification of these speed and accuracy claims remains unavailable at this time.