Skip to content
ReviewRamble

The weekly brief

AI, reviewed, in your inbox.

  • New reviews of tools, models, MCP servers and repos, with the verdict up front
  • Re-checked prices and ratings when something changes
  • New analysis pieces as they are published

Models, agents, MCP servers and the tools actually worth your time.

One email a week. Unsubscribe with the link in any brief. How we use your email

Gamer (miHoYo)

miHoYo's research agent that plays games from the screen with keyboard and mouse. The paper is out; the trained model is not.

Reviewed by ReviewRamble editors · Updated

View on GitHub Open source

Share this page:XLinkedInFacebookRedditWhatsAppEmail

Our rating

3.2/5

Gamer is a serious paper about teaching a model to play the way a person does, attached to a repository that, for now, mostly lets you read about it.

What is Gamer (miHoYo)?

Gamer is a vision-language-action model from miHoYo's Honkai AI R&D team, written with researchers at the National University of Singapore, Tsinghua and three other Chinese universities. It looks at the screen five times a second and answers in keyboard and mouse input: one mouse movement plus six 33 ms groups of held keys for every 200 ms, about 30 control steps a second, drawn from a vocabulary of 74 key names. It is built on Qwen3-VL and Qwen3.5 backbones and trained in three stages: continued pre-training on more than 10,000 hours of gameplay, instruction tuning that teaches it to reason only at decision points and keep a running memory summary, and online reinforcement learning with a method the authors call TA-GRPO.

The repository is the paper's companion: the system prompt template, the action format and parsing rules, and a folder for Streaming-vLLM, the low-latency serving engine, marked coming soon. The paper is arXiv 2611.18397. Checkpoints from the related GameScaling study, from 0.8B to 27B parameters, are on Hugging Face; models trained on Honkai: Star Rail are not released. The first commit was on 14 November 2026 and the last on 16 November.

Is Gamer (miHoYo) worth it?

The interesting claim is not that a model can play a game; as the paper notes, agents beat people at Atari, Minecraft and StarCraft years ago. It is that Gamer does so without being let in through the back door. Those earlier champions read internal game state or used custom action interfaces; Gamer gets the pixels and a keyboard, like everyone else on the sofa. On held-out content from Honkai: Star Rail, miHoYo's own game, the reinforcement-learned model completed 80.0% and 83.6% of the milestones in two side quests, which have 30 and 55, and averaged 49.7% across 40 minigames of seven types. Instruction tuning did most of the lifting; reinforcement learning then added 7.8 and 14.5 points on the quests and 10.5 on the minigames.

The baselines make the point more sharply than the abstract does. On the same minigames the paper scores GPT-6.1-Sol at 17.1%, Gemini-3.8-Flash at 2.0% and the untuned Qwen3-VL-8B at 3.2%. A general model can read a puzzle; acting on it in time is another trade. The paper also shows that speed is part of the skill: policies that do well while the game politely pauses for them degrade sharply when it does not, which is the problem Streaming-vLLM exists to solve. On an H200 with Qwen3.5-35B-A3B, the authors cut a request from 1306.00 ms to 65.12 ms, a 20.1× speedup, though the comparison sets one GPU against four.

The catch: very little here runs. The repository is a README, a licence and a placeholder folder; Streaming-vLLM is marked coming soon, and the paper states that, for data security reasons, no checkpoint trained on Honkai: Star Rail data will be released. What you can download are the GameScaling pre-training checkpoints, which stop before the instruction tuning and reinforcement learning that produced the results above. The headline claim of efficiency comparable to skilled players has, in our reading, no human score beside it in any table. Every number comes from the authors, on a game they make, with rewards checked against engine state no outside lab can reach, and training served 32 game environments from 64 H200 GPUs.

Our take: read it as a well-documented field report from a studio teaching a model to play its own game, with an action format worth borrowing. Anyone hoping to download an agent to grind their dailies will find the door politely shut.

Where it earns its place

  • A PhD student working on vision-language-action agents for games

    Adopt a tested action format instead of designing one from scratch

    Take the published chunk of one mouse displacement and six 33 ms key groups, the 74-token key vocabulary and the parsing rules: keep special tokens, parse only after the last </think> tag, treat anything unparseable as a no-op. Free under Apache-2.0 and usable with any model you fine-tune.

  • A QA lead at a mid-sized game studio

    Judge whether screen-driven agents could run regression tests on quests

    Read the paper's sections on verifiable environments and batched evaluation, then start from a GameScaling checkpoint on Hugging Face. Budget for real hardware: during training the authors served 32 Windows game environments from 64 H200 GPUs.

Before you commit

Setup
Nothing to install from this repository yet. Read the paper, copy the system prompt template and parsing rules from the README, and download GameScaling checkpoints from the THUSI-Lab collection on Hugging Face; Streaming-vLLM is marked coming soon.
Integrations
Qwen3-VL and Qwen3.5 dense and MoE backbones; outputs need serving with special tokens kept; training extends the Verl framework; the agent acts on Windows games through ordinary keyboard and mouse events.
Data and privacy
No service and no data collection: the repository is documentation under Apache-2.0. Honkai: Star Rail data and checkpoints are not released, and game screenshots and videos remain their owners' property.
Limits
Flagship models withheld and the inference engine unreleased; results cover one commercial game plus four public benchmarks; the authors say deeper strategy may outrun the model's reasoning and serving costs must fall before wider use.

What’s good

  • Plays from pixels through keyboard and mouse, with no game API
  • Action format, key names and system prompt fully documented
  • Large gains over general models on held-out minigames
  • GameScaling checkpoints from 0.8B to 27B on Hugging Face

What’s not

  • No runnable code in the repository yet
  • Honkai: Star Rail checkpoints withheld
  • All results from the authors, on their own game
  • Training and serving assume H200-class GPUs

Best for

Researchers building screen-and-keyboard agents who want a documented action format and a training recipe to measure against.

Gamer (miHoYo) alternatives

Gamer (miHoYo): common questions

Is Gamer (miHoYo) worth it?
Gamer is a serious paper about teaching a model to play the way a person does, attached to a repository that, for now, mostly lets you read about it. We rate it 3.2 out of 5.
How much does Gamer (miHoYo) cost?
Free and open source (Apache-2.0) for the code and documentation. GameScaling checkpoints are free on Hugging Face; the Honkai: Star Rail models are not released.
What are the downsides of Gamer (miHoYo)?
No runnable code in the repository yet; honkai: Star Rail checkpoints withheld; all results from the authors, on their own game; training and serving assume H200-class GPUs.
Who is Gamer (miHoYo) best for?
Researchers building screen-and-keyboard agents who want a documented action format and a training recipe to measure against.
Is Gamer (miHoYo) safe for company data?
No service and no data collection: the repository is documentation under Apache-2.0. Honkai: Star Rail data and checkpoints are not released, and game screenshots and videos remain their owners' property.
What are the best Gamer (miHoYo) alternatives?
The closest alternatives we have reviewed are OpenComputer and CUA Agent Kit.

Answers checked Nov 21, 2026.