MiHoYo's agent is playing Honkai: Star Rail
miHoYo's Gamer was built on Qwen3-VL-8B.
On 13th November, miHoYo's Honkai AI R&D Team posted a paper to arXiv describing Gamer, a model that plays Honkai: Star Rail by looking at the screen and sending keyboard and mouse input.
Gamer is built on Alibaba's open-weight Qwen3-VL-8B model and trained in three stages:
- continued pre-training on recorded gameplay paired with keyboard and mouse events,
- supervised fine-tuning on annotated instructions, and then
- reinforcement learning.
Each decision covers roughly 200 milliseconds of play, divided into six steps of about 33 milliseconds, and the model writes out its reasoning only when the situation changes.
The system prompt published in the project's repository opens: "You are a gaming expert, and you are currently playing {GAME_NAME}."
The results were reported on held-out tasks in Star Rail: two side quests containing 30 and 55 milestones, and 40 minigame instances across seven types. The final model reached 80.0% and 83.6% of the side-quest milestones and averaged 49.7% across the minigame types.
Reinforcement learning added 7.8, 14.5 and 10.5 percentage points respectively over the fine-tuned model. The abstract describes "operational efficiency comparable to skilled players." What the tables measure is milestones reached, not quests completed, and success on fewer than half of the minigames.
The reinforcement learning stage is where the infrastructure comes in. The paper explained why commercial games are hard to train against saying, "Honkai: Star Rail depends on a full game client, supporting online services, GPU-accelerated rendering, and complex game-state dependencies."
Its answer is what it calls an engine-instrumented sandbox; "we adapt the game engine to expose task-relevant state transitions and completion conditions to the environment backend." Every task was defined by an initialization procedure, a completion check and a termination rule, so the same objective was run repeatedly from the same starting point.
The paper is careful about where that access sits. "Crucially, all privileged engine access remains within the environment backend. The policy receives only screen observations and task context."
So the claim that Gamer played without touching backend code was true of the model while it played but it is not true of how it was trained. The rewards it learned from were checked against the game's internal state, which only miHoYo, as the game's developer, could read.
Training ran 32 parallel Windows instances of the game, served by 64 Nvidia H200 GPUs, collecting 128 attempts per iteration in groups of eight.
The paper also stated what else the sandbox was used for. The same problems, it says, "extend beyond online RL to large-scale automated evaluation and gameplay testing, where hundreds of executions must start from comparable states and be judged against consistent success criteria."
And: "The same initialization, execution, and verification interfaces also support batched evaluation and automated regression tests without manually resetting or judging each attempt." That is a description of a QA system.
It is also the part an outside lab cannot build. A year earlier, on 12th November 2025, ByteDance Seed published Lumine, an agent built on Qwen2-VL-7B and pre-trained on 1,731 hours of human gameplay.
Trained in miHoYo's Genshin Impact, it completed the five-hour Mondstadt main storyline, and without further fine-tuning, it completed the full five-hour first chapter of Honkai: Star Rail.
So an agent from a different company had already played through Star Rail's opening chapter from pixels. miHoYo's own comparison table lists Lumine's recipe as pre-training and fine-tuning. Gamer adds reinforcement learning, and reinforcement learning is the stage that depends on the instrumented engine. The two papers use different benchmarks and should not be ranked against each other. The difference between them is in what each team had access to.
The training data is described precisely, up to a point. The Star Rail recordings comprise roughly 68,000 sessions and 11,000 hours of video with keyboard and mouse events, collected, in the paper's words, "from multiple sources."
Filtering reduced that to about 3,200 sessions and 1,000 hours for annotation, and about 332 hours of Star Rail play remained in the pre-training corpus. The annotation was done by GPT-5.6-Luna, producing around 1.2 million reasoning annotations and 223,000 memory annotations.
Beyond Star Rail, the paper's opening figure lists more than 40 games, from miHoYo's Genshin Impact and Zenless Zone Zero to other publishers' titles including PUBG, Dota 2, Counter-Strike 2, Fortnite and GTA V.
A companion study by an overlapping team, GameScaling, posted on 20th October, drew its gameplay pool from three public datasets: VPT, D2E and Open-P2P.
On authorship, the paper is credited to the miHoYo Honkai AI R&D Team. The corresponding authors are at miHoYo and the University of Science and Technology of China. Contributors are also affiliated with the National University of Singapore, Tsinghua University, the University of Chinese Academy of Sciences and Xi'an Jiaotong University, and most of those with outside affiliations are marked "work done while interning at miHoYo."
Two numbers indicate the scale of the effort. The scaling study behind the pre-training used approximately 45,000 H200 GPU-hours. And the team reported cutting per-request latency from 1,306 milliseconds to 65.12 milliseconds on a 35-billion-parameter mixture-of-experts model running across four H200s, which is what made real-time play possible at all.
The agent is the reproducible part. ByteDance built one without miHoYo's involvement. The instrumented client is the part that is not reproducible. A publisher that can start any task from a fixed state, run hundreds of attempts in parallel and have the game itself mark the result has a regression suite that plays the game. Gamer is the first detailed account of a publisher building one around its own live title.