Tencent Internal Memo Leaked: Latest Model Hy3 Utilized Claude Code in the Training Phase

According to Data Beating monitoring, The Information revealed, citing sources familiar with the matter and an internal Tencent memo, that Tencent employees used Anthropic's Claude Code in the post-training phase of developing the Hy3 model, even though Anthropic explicitly prohibits providing commercial services to Chinese companies on grounds of national security.

Hy3 is Tencent's most powerful large language model in the Omni series to date, with a 295B parameter MoE architecture, developed under the leadership of Tencent's Chief AI Scientist, Yao Shunyu, who joined the company at the end of last year. During the RLHF (Reinforcement Learning from Human Feedback) phase, Tencent organized employees to act as human evaluators. An internal memo provided installation instructions for the Claude Code, with a per-person usage cap of "thousands of tokens."

Tencent employees do not consider this to be distillation (training a weak model from a strong model's output). Their approach involves posing the same programming problem to two anonymous models for blind grading, while simultaneously using the Claude Code to generate high-quality behavioral examples in real-time as references to help filter out low-quality responses. Employees from multiple AI companies confirmed that using industry-leading models as a benchmark reference during post-training is a common practice.

An Anthropic spokesperson stated that the company's security team "proactively monitors distillation attacks and takes immediate action upon discovery," but did not directly address Tencent's use of the Claude Code. Common ways for Chinese companies and developers to access Claude include using intermediaries, non-Chinese phone numbers, or credit card registrations. Earlier this month, Anthropic tightened its authentication requirements, with some users now needing to provide government-issued ID photos and personal pictures.