GameScaling

GameScaling

A small vision-language model built on Qwen3.5-0.8B from a research project on continued pre-training for embodied agents in many virtual game worlds. For researchers working on game agents.

Made by THUSI-Lab on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: GameScaling is an image text to text model from THUSI-Lab, built on Qwen3.5-0.8B. It belongs to a research project called Scaling in Games, which studies continued pre-training of agents for diverse virtual worlds. The authors come from miHoYo, Tsinghua University and other institutions, and the page lists several sizes with S1 and S3 labels.

Who it suits: Researchers who study embodied agents and game environments and want a small vision-language model to start from.

Industry | Who uses it | How they use it | Example
Game AI | Teams studying game agents | Start from a small vision-language model tuned for virtual worlds | Feed in a game screenshot and an instruction
Embodied AI research | Academic labs | Study continued pre-training across many worlds | Compare the 0.8B and 2B sizes
Machine learning education | Students | Read a small model tied to a published paper | Download it and inspect the files

You can skip it if you want a ready-made chat model, since the card in view gives few run steps.

Where to use it:

• Python with the transformers library
• Windows, macOS or Linux, to download the files
• Research setups for agents in virtual worlds

Use case | Who | What they get back
Image and text prompts | Researchers | A text reply that reads an image plus a question
Size comparison | Experimenters | Results from the 0.8B or 2B variants
Agent research | Game AI teams | A base for agents that must act in virtual worlds

How to set it up: You need Python and the transformers library. The part of the card shown here has no run steps, so follow the model page and the project's code repository for the exact loading code.

• Python and pip
• Enough memory for a small model
• The project's code repository

1. Download the model files.
2. Open the model page and the code repository for loading steps.
3. Load the model with transformers.

pip install -U huggingface_hub
hf download THUSI-Lab/GameScaling

How to use it: You give the model an image and a text prompt, and it returns text. The exact prompt format is not shown in the part of the card used here, so check the model page before you build on it.

Example: You send a screenshot from a game with the question "What should the character do next?". The model replies with text. The card does not show a sample, so the quality of such answers is untested here.

For newcomers and experts: If you are new, download the files and read the model page before you try to run anything. This is a research release, so expect to read some code.

For experts, the base is Qwen3.5-0.8B, and the project name points to scaling the pre-training with game data for embodied agents. The sibling releases come in 0.8B and 2B sizes with S1 and S3 labels. The card shown here does not define those labels.

Game image and text prompt
Qwen3.5-0.8B based model
Text reply

Good to know:

• Base model: Qwen3.5-0.8B
• Language: English
• Variants: 0.8B and 2B sizes with S1 and S3 labels, which the card shown here does not explain
• License: Apache 2.0
• Cost to run: no API key or paid service named
• Data it reaches: downloads files from Hugging Face, nothing else is named
• Not for: the part of the card used here names no limits, and it is a research release

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: transformers
• Downloads in the last 30 days: 635
• Likes: 65
• License: apache-2.0
• Languages: en
• Based on: Qwen/Qwen3.5-0.8B
• Tags: transformers, gamescaling, game-agent, embodied-agent, keyboard-mouse, continued-pretraining, qwen3_5, image-text-to-text

Links

• Model: https://huggingface.co/THUSI-Lab/GameScaling
• Files: https://huggingface.co/THUSI-Lab/GameScaling/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download THUSI-Lab/GameScaling

Then follow the usage in its model card. Nothing in this kit runs by itself.