# Gamer VLA Model: miHoYo's Long-Horizon Game Agent Explained

> Source: <https://explainx.ai/blog/mihoyo-gamer-vla-foundation-model-long-horizon-game-agent-2026?utm_source=rss&utm_medium=feed&utm_campaign=blog>
> Published: 2026-11-16 00:00:00+00:00

**Gamer is miHoYo's new [vision-language-action (VLA)](https://explainx.ai/dictionary/vision-language-action-model-vla) foundation model that plays long 3D games from screen pixels and human-style keyboard and mouse commands — no hidden game-state API.** The Honkai AI R&D team posted the [arXiv paper](https://arxiv.org/abs/2611.18397?ref=explainx) on **November 13, 2026**, opened the [project page](https://mihoyo-gamer.github.io/?ref=explainx), and shipped [Apache-2.0 code](https://github.com/RainBowLuoCS/Gamer?ref=explainx) that documents how a single policy reasons, remembers, and acts across **hour-long quests**.

For builders, Gamer is not a miHoYo storefront feature. It is a **research stack** — continued pre-training on **tens of thousands of hours** of gameplay, **sparse-thinking** supervision, **Temporal-Aware GRPO (TA-GRPO)** reinforcement learning, and **Streaming-vLLM** serving aimed at **sub-100 ms** decision loops. That combination is why the Hugging Face Daily Papers entry hit **#2** with rising upvotes: it connects [embodied agent roadmaps](https://explainx.ai/blog/envisioning-agentic-robotics-survey-s0-s1-s2-roadmap-2026) to something millions of players already understand — **beating a hard RPG quest under time pressure**.

## TL;DR

| Question | Answer | 
|---|---|
| Who released it? | **miHoYo Honkai AI R&D Team** (authors on arXiv and GitHub) | 
| What interface? | **Screen observations** in,**keyboard and mouse** action tokens out | 
| Training recipe? | **CPT** on massive trajectories →**sparse-reasoning SFT** →**TA-GRPO** online RL | 
| Serving? | **Streaming-vLLM** with KV reuse and speculative decoding; README cites**~20×** latency drop on H200/Qwen3.5-35B-A3B MoE | 
| Code? | [GitHub RainBowLuoCS/Gamer](https://github.com/RainBowLuoCS/Gamer?ref=explainx) — prompts, parsers, benchmark tables; weights not bundled as a product drop | 
| How does this compare to cheat headlines? | Unlike [GPT-6 Astra's StarCraft shortcut](https://explainx.ai/blog/gpt-6-astra-starcraft-starskirmish-downloads-human-bot-cheat-2026) , Gamer's setup forbids swapping in human bots mid-episode; rewards come from**environment verification** | 

## Why miHoYo built a game-native VLA now

Games are **stress tests** for agents: partial observability, **long horizons**, sparse rewards, and **real-time control** at 30–60 FPS. Productivity [computer-use agents](https://explainx.ai/blog/alibaba-opengcua-open-stack-generalist-computer-use-agents-osworld-v2-2026) often click through forms. RPGs demand **memory** (“which door did I already try?”), **combat sequencing**, and **navigation** across maps that do not fit one context window.

miHoYo's angle in the [abstract](https://arxiv.org/abs/2611.18397?ref=explainx) is **generalist control**: the **same action vocabulary** (mouse deltas, key holds, scroll) should work in **2D benchmarks** and **3D open worlds**, and trajectories can mix **virtual games** with **embodied simulators** so dynamics transfer. That mirrors how [Microsoft's idle-GPU gaming patent](https://explainx.ai/blog/microsoft-patent-idle-gaming-gpu-micro-training-us-20260349820-2026) and [Xbox Labs' AI porting experiments](https://explainx.ai/blog/xbox-labs-ai-game-decompiling-recompiling-ports-asha-sharma-2026) treat commercial titles as **data factories** — except Gamer publishes the **training and inference recipe**, not just IP around it.

**What games did miHoYo make?** The studio is known for **Genshin Impact**, **Honkai: Star Rail**, **Zenless Zone Zero**, and related live-service titles. **Gamer is not a new game in that list**; it is an **R&D agent** evaluated partly inside **Honkai: Star Rail** quests, with demo art on the project page referencing multiple franchises for **benchmark variety**.

## Architecture: perception, sparse reasoning, and memory

The policy is end-to-end: **vision encoder + language backbone + action head** that emits structured keyboard and mouse chunks. Three ideas separate it from a vanilla VLM fine-tune:

1. **Unified visual control** — actions are**device-level** trajectories aligned to frames, not abstract joystick vectors tied to one title.
2. **Sparse reasoning** — the model can output**action-only** tokens when the prior plan still works, and only opens a reasoning block when objectives change (documented in the[GitHub system prompt template](https://github.com/RainBowLuoCS/Gamer?ref=explainx) ).
3. **Persistent memory summaries** — when the context window fills, the agent writes**compact memory** of finished subgoals, failures, and remaining plan steps, then continues rolling forward.

## Sparse reasoning and cross-context memory

Gamer VLA model sparse reasoning and cross-context memory architecture diagram
*Image: [miHoYo Honkai AI R&D Team](https://mihoyo-gamer.github.io/?ref=explainx)*

The figure above (from the official project page) shows how **reasoning** and **memory updates** share the **same autoregressive stream** as low-level actions. That matters for deployment: you do not run a separate planner service that drifts from the executor.

Practically, each decision covers about **200 ms** split into **six ~33 ms steps**. Mouse movement applies to the whole chunk; keys can differ per step. The README lists **74 case-sensitive key tokens** and strict parsing rules — including treating malformed outputs as **no-ops**, which is how you keep an RL farm from melting down when the policy hallucinates format.

If you build agents, steal the pattern: **one tokenizer for thought, memory, and act**, with the harness enforcing **FSM-shaped structure** so speculative decoding can propose safe delimiters.

## Training stages: CPT, SFT, and TA-GRPO

The [GitHub method table](https://github.com/RainBowLuoCS/Gamer?ref=explainx) breaks training into three supervised regimes plus online RL:

| Stage | Supervision | What the policy learns | 
|---|---|---|
| **Continued pre-training (CPT)** | Screen frames paired with keyboard and mouse from diverse games plus general multimodal data | Transferable visual control and dynamics before task-specific rewards | 
| **Sparse-reasoning SFT** | Multi-level instructions with **~1.2M reasoning** and**~223K memory** annotations (per Hugging Face team post) | When to think, what to remember, and how to finish subgoals | 
| **TA-GRPO** | Online rollouts with **verifiable environment outcomes** | Long-horizon credit assignment across segments linked by memory | 

**Temporal-Aware GRPO** is the paper's answer to **context limits** and **rollout error accumulation**: instead of pretending an hour-long episode fits one forward pass, training stitches **memory-linked segments** and assigns reward along the timeline. That is the game analogue of [multi-turn agent RL](https://explainx.ai/blog/uni-robot-agent-open-qwen-robodojo) harnesses in robotics — except rewards are **quest milestones** and **minigame win conditions** you can check automatically.

The team also continues a parallel **GameScaling** CPT line ([arXiv:2610.28698](https://arxiv.org/abs/2610.28698?ref=explainx), cited in the repo) on **scaling laws in diverse virtual worlds**, which explains why they emphasize **10K+ hours** of trajectory CPT in public messaging.

Gamer VLA multi-stage gameplay data curation and annotation pipeline
*Image: [miHoYo Honkai AI R&D Team](https://mihoyo-gamer.github.io/?ref=explainx)*

## Honkai: Star Rail results the README publishes

Gamer foundation model Honkai Star Rail side quest benchmark results chart
*Image: [miHoYo Honkai AI R&D Team](https://mihoyo-gamer.github.io/?ref=explainx)*

On held-out **Honkai: Star Rail** tasks, the README posts concrete **milestone completion rates** after each stage. Side quests **Opportunity Makes the Thief (Part I)** and **Bestial Ferocity** climb from **34.4% / 47.9%** after CPT to **80.0% / 83.6%** after RL. **Minigame macro-averages** rise from **15.0%** to **49.7%**.

Those numbers are **internal eval tables**, not a promise that a retail copy of Star Rail will ship a autopilot mode. They do show **TA-GRPO adds double-digit points** beyond SFT — the kind of lift you want before trusting an agent on **55-milestone** quests.

The Hugging Face announcement further claims **over 95%** success on **main-storyline challenges** in large-scale automated tests. Treat that as **author-reported coverage** until independent replications appear; still, it signals the model is meant for **production-scale QA automation**, not couch play.

## Streaming-vLLM: why inference speed becomes a training variable

Long-horizon RL fails if **action latency** blows the game's tick budget. Gamer's serving stack, **Streaming-vLLM**, targets **concurrent online rollouts** with:

- **Per-environment KV-cache reuse** and memory compaction
- **Verified finite-state-machine speculative decoding** for structural tokens
- **Early action dispatch** so the client executes the first valid chunk while later tokens stream
- **Architecture-specific paths** for**Qwen3-VL** and**Qwen3.5** dense and MoE backbones

The README's headline table compares **1306.00 ms** baseline latency to **65.12 ms** after cumulative optimizations on **NVIDIA H200** with **Qwen3.5-35B-A3B MoE** and **TP4** — about **20.1×** faster. The Hugging Face comment rounds this to **20×** and argues faster inference **improves online RL**, not just player feel.

Implementation is marked **coming soon** in the `streaming-vllm/` directory. Until it lands, treat the engine as a **paper plus design doc**, not something you can npm install for Friday's demo.

## How to try Gamer as a builder

1. **Read the paper** —[arXiv:2611.18397](https://arxiv.org/abs/2611.18397?ref=explainx) for TA-GRPO and data-curation figures.
2. **Clone the repo** —[github.com/RainBowLuoCS/Gamer](https://github.com/RainBowLuoCS/Gamer?ref=explainx) for system prompts, action grammar, and benchmark tables.
3. **Watch the project page demos** —[mihoyo-gamer.github.io](https://mihoyo-gamer.github.io/?ref=explainx) hosts gameplay films (third-party game assets remain their owners' IP; Apache-2.0 covers code, not art).
4. **Reproduce parsing, not full quests** — start by implementing the**action parser** and**sparse reasoning gating** on a toy 2D env before attaching a 35B MoE.

There is **no Hugging Face model card with open weights** attached to the November drop. The repo's roadmap still tracks **Streaming-vLLM code**, extra demos, and documentation. Plan for **research access**, not a Battle.net-style bot.

### Action format cheat sheet

Each chunk looks like:

`<|action_start|>X Y Z ; g1 ; g2 ; g3 ; g4 ; g5 ; g6<|action_end|>`

**X** and **Y** are mouse deltas in thousandths of screen width and height (clipped to ±1000). **Z** is scroll. **g1…g6** are key groups per 33 ms step. Reasoning, when needed, lives in a `<think>` block immediately before the action block — the server must parse **only the final block after the last closing tag** so it does not execute examples nested inside thoughts ([README parsing section](https://github.com/RainBowLuoCS/Gamer?ref=explainx)).

## Benchmarks beyond one RPG

Project-page art references evaluations across **Genshin Impact**, **Zenless Zone Zero**, **Minecraft**, **DOOM Eternal**, and classic **2D** suites — a spread designed to show **interface generalization**, not overfitting one UI skin. The paper abstract claims **competitive 2D benchmark scores** while emphasizing **3D open-world generalization**.

That positioning sits next to [WorldFoundry's multi-benchmark orchestration](https://explainx.ai/blog/worldfoundry-open-world-model-benchmark-infrastructure-2026) trend: games become **standardized agent gyms** when you can verify rewards programmatically.

## Safety, cheating, and live-service risk

Gamer's research story lands next to Korea's **generative-AI exam cheating** scandal and everyday **anti-cheat** arms races. Technically, a screen-and-mouse policy is **indistinguishable from a skilled human** at the input layer — which is why online games rely on **server authority**, **behavioral stats**, and **kernel anti-cheat**, not “please don't use AI.”

For builders shipping agents into **live games**:

- Assume **Terms of Service** ban unsanctioned automation even when inference is local.
- Prefer **offline simulators** or**first-party QA environments** with signed clients.
- Log **provenance** of trajectories if you fine-tune further — publishers increasingly audit**synthetic play data** .

This is parallel to why explainx.ai covered [Astra's StarCraft rule break](https://explainx.ai/blog/gpt-6-astra-starcraft-starskirmish-downloads-human-bot-cheat-2026) as a **harness failure**, not a skill flex.

## What we still do not know

- **Weight release scope** — paper and code are public;**checkpoint hosting** is not guaranteed.
- **Streaming-vLLM open-source date** — README says**coming soon** .
- **Independent replication** of**95%+ storyline** claims on consumer hardware.
- **Latency numbers** outside the quoted**H200 / Qwen3.5-35B-A3B MoE** configuration.

Watch the [GitHub releases](https://github.com/RainBowLuoCS/Gamer?ref=explainx) and the [Hugging Face paper thread](https://huggingface.co/papers/2611.18397?ref=explainx) for weights, not rumor accounts.

## Related reading

Reading gets you the idea. Building it with feedback gets you the skill.
