腾讯混元

腾讯混元联合多所高校推出MMAE，首个全面评估AI音频编辑能力的基准，揭示当前模型在精确编辑音频任务上的不足。

Tencent HunYuan releases the open-source Hy-MT2 translation model supporting 33 languages. The 7B and 30B-A3B models achieve top performance among open-source models, while the 1.8B model outperforms commercial APIs. With 1.25-bit quantization, it requires only 440MB storage for mobile deployment.

Tencent's AI 'Second Half' strategy centers on Coding Agents as a foundation and leveraging small models for high-value tasks. Shunyu Yao emphasizes that real-world data from Yuanbao DAU is more critical than benchmarks for fixing model baseline issues and building context barriers.

Tencent Hunyuan team releases open-source multilingual translation model Hy-MT2, supporting 33 languages, with 7B and 30B-A3B versions achieving SOTA performance among open-source models, the 1.8B lightweight version even surpasses mainstream commercial APIs, using 1.25-bit extreme quantization technology requiring only 440MB storage space.

Tencent Hunyuan Hy3 preview launched, designed for complex agentic tasks using rebuilt pre-training and reinforcement learning infrastructure, outperforming benchmark-focused models in real-world effectiveness.

Tsinghua and Tencent Hunyuan team won the championship in the MLSys 2026 MoE model inference optimization competition, showcasing efficient inference optimization techniques.

Within two weeks of its launch, Hy3 preview saw a 10x increase in token calls, but the article lacks in-depth analysis and specific technical details.

Tencent launches a lightweight translation large model that can run locally on mobile phones without internet.