Tencent releases and open-sources Hy3 model, rivaling larger flagship models

Tencent's Hunyuan team today released and open-sourced the new large language model Hy3. Based on the preview version from late April, the model incorporates feedback fixes and experience improvements, focusing on enhancing the quality and diversity of post-training data while expanding reinforcement learning computing power.

Hy3 shows significant progress in tasks like reasoning, agent, and long context, with claims that its performance rivals larger flagship models two to five times its size. It offers high cost-effectiveness in productivity tasks such as software development, office work, financial modeling, front-end design, and game creation.

In a blind test involving 270 internal experts working on real tasks, Hy3 scored 2.67 out of 4, outperforming GLM5.1's 2.51, especially in front-end, data storage, and CI/CD categories. Tencent also reduced usage costs for Hy3 on its cloud API and products like WorkBuddy through software-hardware optimization, with input pricing at 1 yuan per million tokens, output at 4 yuan, and cache hits at 0.25 yuan.