miHoYo Unveils Gamer, a Generalist VLA Game Agent
HoYoverse has released Gamer, a general-purpose game agent built as a vision-language-action (VLA) foundation model for long-horizon real-time games: it watches screen screenshots directly (sampled at 5 Hz) and outputs keyboard and mouse actions, relying on no game APIs at all, and can play a variety of games in real time like a human. The project is published on arXiv and open-sourced.
Confirmed
- The architecture builds on a pretrained multimodal model, taking 5 Hz screenshots as input and outputting a unified keyboard/mouse action format: mouse displacement plus six consecutive 33ms key states every 200ms (as minchoi relayed from the project page)
- Training proceeds in three stages: continued pretraining (CPT) on over 10,000 hours of device-level keyboard/mouse interaction traces covering both 2D GUIs and 3D simulated worlds, followed by stages including TA-GRPO reinforcement learning
- A single unified model covers more than 30 games, including Genshin Impact, Honkai: Star Rail, Zenless Zone Zero, PUBG, Minecraft, and GTA
- It achieves an over 95% clear rate on Honkai: Star Rail challenge quests; demos span navigation, dialogue, combat, UI operations, and puzzle-solving, with the project page showcasing 7 mechanic categories, 32 mini-game recordings, and two long-horizon full completions
- According to Run Luo, the paper comes from HoYoverse's Honkai AI research team, with a framework integrating perception, memory management, and multi-step reasoning; pmttyji noted participation from Tsinghua's THUSI-Lab and other institutions
Why it matters
- The API-free, vision-only design means the approach could in principle generalize to any game operable via keyboard and mouse, and even to broader GUI tasks
- Long-horizon real-time decision-making (second-level action outputs, completions lasting tens of minutes) has long been a challenge for game agents; Gamer demonstrates the viability of the VLA foundation model route in this setting
2026-11-16 ~ 2026-11-17 · 9 related posts
Primary sources
- [source] miHoYo unveils Gamer, a VLA foundation model for real-time game agents — Run Luo · 2026-11-16
- Gamer: A Generalist VLA Foundation Model Trained on Tens of Thousands of Hours of Keyboard-Mouse Gameplay — pmttyji · 2026-11-16
- miHoYo's Gamer agent beats 95% of Honkai: Star Rail challenges, plays 30+ games — _akhaliq · 2026-11-16
- miHoYo unveils GAMER, a generalist game agent playing 32 minigames with one interface — BinarySplit · 2026-11-17
- [source] miHoYo unveils Gamer, a generalist AI agent that plays games by watching the screen — minchoi · 2026-11-17
- miHoYo unveils Gamer, an AI that plays games by watching the screen with no API — minchoi · 2026-11-17
- Gamer AI aims and fires in a 3D shooter it never trained on — minchoi · 2026-11-17
- Gamer AI plays through Doom straight from the pixels — minchoi · 2026-11-17
- [source] Gamer technical report: 10K hours of gameplay, CPT+SFT+TA-GRPO RL, Streaming-vLLM — minchoi · 2026-11-17