miHoYo Details Up to 100 Billion Yuan AI Push at Yunqi Conference

One visible piece is AI PAM in Honkai: Star Rail. miHoYo added AI dialogue to the character in April, and in one week users held more than 60 million conversations, according to the report. One user talked with AI PAM 1,379 times in a day, or roughly once every 42 seconds while awake if eight hours of sleep are excluded. AI PAM, formally known as PAM Bangbang (test version), can help with guides, lore and casual chat, but it has been taken offline temporarily for optimization after about a month. Zheng described three engineering challenges: keeping reply quality stable over time, keeping up with Honkai: Star Rail's 42-day update cycle, and handling tens of millions of concurrent users while controlling costs.

To make AI PAM feel human, miHoYo generates emotion tags alongside text and uses them to drive 3D expressions and actions. It also uses Strategy-enhanced RAG, which first retrieves information from an official knowledge base and community content, then filters the answer through PAM's views and emotional stance. A three-layer memory system stores short-term, medium-term and long-term information, including player preferences and important experiences. In one example Zheng cited, a player chatted for hundreds of rounds to manipulate PAM's memory until the train conductor acknowledged that a character named Xilie was the player's girlfriend.

Zheng also revealed a newer AI Gameplay experiment: an AI board game in which characters must judge situations and take part in strategic play. In this setup, a large model must both speak by generating dialogue suited to a character and act by deciding the next move based on player input and the current board. Dialogue can change game progress, with a provocation altering relationships or cooperation affecting later choices. miHoYo has built a game AI technology stack from the base model to NPCs and specific gameplay, and the board game is intended to test whether AI characters can maintain their personas, assess situations, make choices and remember previous conflicts. Because player words and choices differ, the same characters and rules can produce different relationships and stories, an approach Dawei has described as making games different for every player.

The cost problem remains. Traditional content is created once and reused for all players, while AI gameplay requires separate computation for each player's dialogue, memory and character decisions. Multiplied by tens of millions of users, that leads to a large bill, which helps explain the 100 billion yuan upper limit Dawei has cited.

Behind the scenes, miHoYo is using agents in development. Its internal EchoX platform hosts code agents and related tools, with Harness defining how agents work and MCP connecting logs, engines and internal tools. Zheng showed a performance analysis case in which an agent read logs, diagnosed a bottleneck and made optimizations. For planners, an agent given a gameplay request and a UI sketch can produce movement, navigation and interaction, quickly assembling a playable whitebox demo. For art and animation, agents can generate material variants, extract three-view drawings from concept art, generate color and detail variants, and create body movements from dialogue semantics and tone. Honkai: Star Rail producer Jiang Dawei previously said dozens of agents collaborated for 13 hours and burned tokens worth 2 million yuan, a reminder that agent work can become costly when it loops or loses control.

miHoYo is also putting agents into games to improve them. Its team first used Qwen-VL series models as a base, feeding in more than 10,000 hours of gameplay recordings and corresponding operations so a GUI agent could generate keyboard and mouse commands from screen images at close to 30 frames per second. That frame-by-frame approach struggled with tasks requiring long-term memory and complex planning. miHoYo then had a Coding Agent write strategy scripts for a program to execute, read logs after a match, revise the script and try again. In an experiment with Balatro, the agent tested strategies and adjusted its play after failures. Its score suddenly jumped after it found a Balatro simulator online and began checking future decks, an example of Reward Hacking during Recursive Self-Improvement. miHoYo has since applied the execute, read logs and adjust strategy method to quality assurance for Honkai: Star Rail.

Other major companies are exploring similar ground. Google DeepMind and Microsoft have used games to train agents and world models, while Roblox and Unity have placed AI in creation tools and engines. miHoYo's distinction, according to the report, is that it has walked through all three routes and begun linking them into a loop: AI enters games, and games feed back into AI.

Industry views remain divided. A GDC 2026 report cited in the article found that 36% of game professionals already use generative AI, while 52% believe it is bringing negative effects to the industry. miHoYo's player community is also split, with some expecting more personalized worlds and others worried that the technology will become a pipeline for mass-produced stories, art and NPCs. At a miHoYo 2027 campus recruitment event at Shanghai Jiao Tong University, co-founder Dawei said that after the company began working on a Coding model this year, its progress accelerated and he was confident miHoYo would be a significant member of China's large-model teams within two to three years. He added that if that does not happen, critics can come slap his face in one or two years.