Tencent internal memo leaked: The latest model Hy3 used Claude Code during the post-training phase

According to Beating Monitoring, The Information cited insiders and internal Tencent memos revealing that Tencent employees used Anthropic’s Claude Code during the post-training phase of developing the Hy3 model, despite Anthropic explicitly prohibiting providing commercial services to Chinese companies on national security grounds.

Hy3 is Tencent’s strongest large language model in the Mengyuan series to date, with 295 billion parameters and a MoE architecture, led by Tencent’s Chief AI Scientist Yao Shunyu, who joined at the end of last year. During the RLHF (Reinforcement Learning with Human Feedback) stage, Tencent organized employees to act as human evaluators, with internal memos providing installation instructions for Claude Code, with a usage cap of “several thousand tokens” per person.

Tencent employees do not consider this to be distillation (training a weaker model using the output of a stronger model). Their approach is to present the same programming problem to two anonymous models for blind scoring; simultaneously, they use Claude Code to generate real-time high-quality behavior examples as references to help filter out low-quality responses. Employees from multiple AI companies confirmed that using industry-leading models as benchmarks during post-training is a common practice.

An Anthropic spokesperson stated that the company’s security team “actively monitors distillation attacks and takes immediate action upon discovery,” but did not directly respond to Tencent’s use of Claude Code. Chinese companies and developers typically access Claude through intermediaries, non-Chinese phone numbers, or credit card registration. Anthropic has tightened identity verification requirements earlier this month, requiring some users to provide government-issued ID photos and personal photos.