miHoYo Unveils Gamer, a General-Purpose Vision-Language-Action Foundation Model That Can Play *Honkai: Star Rail* and Other Games Autonomously
ITHome, November 17 — The AI R&D team behind miHoYo's Honkai series has unveiled "Gamer," a general-purpose vision-language-action foundation model. The model can autonomously play 3D and 2D games by simulating how a real person operates them.

According to information published in the research paper and on the project page, Gamer does not read internal game data and does not rely on game-specific APIs. Instead, it captures game frames at 5 frames per second (5 Hz) as input, then generates native keyboard key states and mouse movement commands equivalent to 30 Hz, thereby simulating the process of a real person playing a game with a keyboard and mouse.

To address the memory constraints and inference latency faced during long-duration task execution, the research team conducted continual pretraining on more than 10,000 hours of cross-game keyboard-and-mouse operation trajectories, and combined sparse thinking instruction fine-tuning with temporal-aware reinforcement learning to optimize the model. In addition, the team developed a dedicated inference engine called "Streaming-vLLM." On NVIDIA H200 hardware, this engine reduced end-to-end request latency from 1306 milliseconds to 65.12 milliseconds.
The research team also demonstrated Gamer carrying out main story missions, exploring scenes, and completing puzzle challenges in Honkai: Star Rail in live gameplay. Test data shows that Gamer-RL, after reinforcement learning training, achieved an average success rate of 49.7% across 40 previously unseen mini-game tasks, outperforming the general model without fine-tuning.

In response to concerns that the technology could be used to develop game cheats and automation scripts, the research team emphasized that Gamer is an academic research project aimed at exploring general intelligent agents for interactive environments, with potential applications including game testing, accessibility assistance, and human-computer interaction research. The team explicitly stated that it does not support using Gamer for game cheating, unauthorized access, exploitation of security vulnerabilities, or any behavior that harms user interests or undermines game fairness.
For data security reasons, the miHoYo research team has currently only released the general files used in the early foundational pretraining research phase of the project. It has not released any Honkai: Star Rail-related training data, nor has it released model weights trained or fine-tuned on that game's data.

Related Articles
- Individual who illegally obtained and cracked Honkai: Star Rail beta content ordered to pay miHoYo 1 million yuan
- miHoYo launches Honkai: Star Rail Huohuo Huohua series merchandise accessories, including a light-up socket, data cable, GaN charger, and more
- Shanghai High People's Court reports that a new version of Honkai: Star Rail was "unpacked and leaked" by a minor; miHoYo awarded 1 million yuan
- miHoYo's Honkai: Star Rail Xilian custom mechanical keyboard opens for pre-order: 108 keys, tri-mode, dual switch options, 808 yuan
- miHoYo & Samsung collaborate again: Galaxy S26 Ultra Honkai: Star Rail Xilian accessory gift box to launch in South Korea on April 24
- miHoYo announces that Honkai: Star Rail version 4.0 raises device performance requirements