GameDojo-4B
AI Agent for Playing Video Games
1 min read
GameDojo-4B is a specialized Vision-Language Model (VLM) trained to control characters in computer games. It analyzes the game screen and task context, outputting a sequence of keyboard and mouse commands. Each generated action block covers 200 ms of gameplay. Created by researchers based on the Qwen3.5-4B architecture, the model was trained on a balanced dataset of game sessions in 2D and 3D environments.
Key features
- Real control: Generates precise mouse and keyboard commands for complex gaming situations.
- Base model: Built on the efficient Qwen3.5-4B language model with image support.
- Wide coverage: Trained on 120 game boss encounters of varying difficulty.
- BF16 format: Distributed in Safetensors weights with standard library support.
What you can do
- Agent testing: Run the model as an autonomous player in supported GameDojo environments.
- CUA research: Explore computer-use capabilities in dynamic simulations.
- Transformers integration: Plug the checkpoint into custom action generation pipelines via Hugging Face
transformers.
Example
import torch
from transformers import AutoProcessor, Qwen3_5ForConditionalGeneration
repo_id = "THUSI-Lab/GameDojo-4B"
processor = AutoProcessor.from_pretrained(repo_id)
model = Qwen3_5ForConditionalGeneration.from_pretrained(
repo_id, dtype=torch.bfloat16, device_map="auto"
).eval()Read more
DetailsBase model: Qwen/Qwen3.5-4B
Parameters: 4.54 billion
Training: Supervised fine-tuning on 240 hours of GameDojo-Data gameplay
Weights: BF16 Safetensors










Comments