What is NitroGen?
NitroGen is an open foundation model for game-playing agents, released by an NVIDIA-led research team in December 2025, with the paper following on 4 January 2026. It looks at the game screen and outputs controller input: it takes 256 by 256 RGB frames and returns two analogue-stick vectors and 17 button presses. It does not generate games; it plays them.
The model has 493 million parameters: a SigLip2 vision transformer feeding a diffusion transformer that produces actions. It was trained by behaviour cloning on 40,000 hours of gameplay from more than 1,000 games, using public videos in which streamers showed their controller inputs on screen. The team also released that dataset, an evaluation suite and a wrapper that exposes any Windows game through the Gymnasium API.
The weights are under NVIDIA's One-Way Noncommercial licence; the SigLip2 backbone is Apache 2.0.
Is NitroGen worth it?
NitroGen is an open game-playing agent you can actually run: one 493M-parameter model that plays action games, platformers and racers from the screen alone, with no game-specific code, and it ships with its weights, code and training data.
The data is what makes it work. The team turned 71,000 hours of streamed gameplay into 40,000 hours of frame-and-button pairs from 38,739 videos by 818 creators, by reading the controller overlays streamers put on screen. On games it never saw in training, the paper reports up to 52% higher task success than models trained from scratch, which is the number that matters if you want to fine-tune it on your own game.
The catch: coverage is uneven and so is skill. Only 91 of the 1,000-plus games have more than 100 hours of data, and 15 have over 1,000, so it is much better at popular titles. The model card says it is weak on mouse-and-keyboard games such as strategy and MOBA titles, and the paper's headline is relative improvement, not a claim that it finishes games. In the paper's evaluation the game is paused while the model picks each action; real-time play is left to future work. The play script attaches to a running Windows game process and sends controller input on your machine, so run it on a test PC, not one with accounts you care about. Commercial use is not allowed.
Adoption is real for a research release: 2,237 GitHub stars and 579 likes on Hugging Face. The GitHub repository's last push was in January 2026.
Our take: for a studio it is a preview of automated play-testing to evaluate, not a tool to deploy, because the licence does not cover commercial work; for researchers, the dataset is worth more than the model.
Where it earns its place
A research engineer in a game studio's R&D group
Find out whether a general agent could one day play-test controller games
Serve the checkpoint with scripts/serve.py and attach scripts/play.py to the build's executable on a dedicated Windows test machine; the paper's unseen-game results suggest fine-tuning on the team's own play helps most. The licence covers non-commercial evaluation only, not production QA.
A PhD student in imitation learning
Study cross-game transfer without building a dataset
Use the released 40,000-hour dataset and evaluation suite as a fixed benchmark, and the Gymnasium wrapper to run experiments on commercial games.
Before you commit
- Setup
- Clone MineDojo/NitroGen, pip install -e ., download ng.pt from nvidia/NitroGen, start the inference server, then run the play script against a Windows game executable.
- Integrations
- Gymnasium-style wrapper for commercial Windows games; Hugging Face dataset nvidia/NitroGen.
- Data and privacy
- Runs locally. The agent reads your screen and injects gamepad input into the target process. Weights under NVIDIA's One-Way Noncommercial licence.
- Limits
- 256x256 input; gamepad actions only; poor on RTS and MOBA games; Windows games only; evaluated with the game paused between actions; non-commercial.
What’s good
- Plays real Windows games from screen pixels to gamepad input
- Dataset, evaluation suite and weights all released
- Up to 52% better than training from scratch on unseen games
- Small at 493M parameters
What’s not
- Non-commercial licence
- Weak on mouse-and-keyboard genres
- Training data concentrated on popular games
- Evaluated with the game paused between actions, not in real time
Best for
Game-AI researchers, and studio R&D teams evaluating automated play-testing, who can work under a non-commercial licence.
NitroGen alternatives
NitroGen: common questions
- Is NitroGen worth it?
- NitroGen is a game agent you can download today and point at a real Windows game, and its non-commercial licence keeps it in the lab. We rate it 3.8 out of 5.
- How much does NitroGen cost?
- Free weights, code and dataset for non-commercial use only. You pay for the GPU.
- What are the downsides of NitroGen?
- Non-commercial licence; weak on mouse-and-keyboard genres; training data concentrated on popular games; evaluated with the game paused between actions, not in real time.
- Who is NitroGen best for?
- Game-AI researchers, and studio R&D teams evaluating automated play-testing, who can work under a non-commercial licence.
- Is NitroGen safe for company data?
- Runs locally. The agent reads your screen and injects gamepad input into the target process. Weights under NVIDIA's One-Way Noncommercial licence.
- What are the best NitroGen alternatives?
- The closest alternatives we have reviewed are Genie 4, Matrix-Game 4.0 and LingBot-World 2.0 (Infinity).
Answers checked Nov 7, 2026.