Tencent Internal Memo Revealed: Latest Model Hy3 Used Claude Code in Post-Training Phase

According to monitoring by Dongcha Beating, The Information cited sources familiar with the matter and an internal Tencent memo revealing that Tencent employees used Anthropic’s Claude Code during the post-training phase of developing the Hy3 model, despite Anthropic explicitly prohibiting commercial services to Chinese companies on national security grounds. Hy3 is Tencent’s most powerful large language model to date, featuring a 295B parameter MoE architecture, and was developed under the leadership of Chief AI Scientist Yao Shunyu, who joined Tencent at the end of last year. During the RLHF (Reinforcement Learning from Human Feedback) phase, Tencent organized employees to act as human evaluators, and the internal memo provided installation guidelines for Claude Code, with a usage limit of ‘thousands of tokens’ per person. Tencent employees did not consider this to be distillation (training a weaker model using the outputs of a stronger model). Their approach involved posing the same programming questions to two anonymous models and conducting blind evaluations; simultaneously, they used Claude Code to generate high-quality behavioral examples in real-time as references to help filter out low-quality responses. Employees from several AI companies confirmed that using industry-leading models as benchmark references during post-training is a common practice. An Anthropic spokesperson stated that the company’s security team ‘actively monitors for distillation attacks and takes immediate action upon discovery,’ but did not directly address Tencent’s use of Claude Code. A common way for Chinese companies and developers to access Claude is through intermediaries or by registering with non-Chinese phone numbers or credit cards. Earlier this month, Anthropic tightened its verification requirements, with some users needing to provide government-issued ID photos and personal photos.