Tencent’s (0700) Hunyuan officially released Hy3 on Monday, and it has been adopted across Tencent's products and services, including WorkBuddy/CodeBuddy, Yuanbao, Marvis, ima, and others.
The company said the feedback from its preview version, launched on April 23 as the first model in the Hy3 generation, led to significant improvements over Hy2 in complex reasoning, instruction-following, in-context learning, code generation, and agent capabilities.
Hy3 has been improved through extensive use by developers worldwide and in Tencent's large-scale real-world business cases. Since the Hy3 preview launched, its average daily token usage has grown twentyfold, indicating increasing market acknowledgment of its practicality and cost-efficiency.
Building on the foundation of Hy3 preview, the official Hy3 delivers enhanced performance through improved post-training data quality and diversity, as well as scaled-up reinforcement learning (RL) compute. It has made significant progress across tasks such as reasoning, agent capabilities, and long-context understanding, achieving performance comparable to flagship models, which often have two to five times the parameter scale of Hy3, both domestically and internationally.
Particularly notable progress has been made in productivity tasks such as software development, office productivity, financial modeling, front-end design, and game production, making Hy3 a reliable and cost-efficient choice. Hy3 is now available under the Apache 2.0 license on open-source communities including GitHub, Hugging Face, ModelScope, and AtomGit.
Pricing is set at 1 yuan (HK$1.16) per million input tokens and 4 yuan per million output tokens, with a cache hit input price of 0.25 yuan. Its API is now available on Tencent Cloud TokenHub, with multiple global third-party developer platforms set to integrate Hy3 progressively
