Laya clears Super Mario Bros. 3 World 1-1 on a DGX Spark: 207 decisions, frames verified

# 🍄 Laya plays Super Mario Bros. 3

**COURSE CLEAR. 207 decisions. Every frame verified.**

[](pyproject.toml)
[](https://github.com/NandhaKishorM/laya)
[](docs/mario.md)
[](https://github.com/cv/laya-plays-smb3/releases/tag/v0.1.0-world1-clear)

[Laya](https://github.com/NandhaKishorM/laya) cleared Super Mario Bros. 3 World 1-1 on a DGX Spark.
**27.50 seconds of game time · 16.79 ms median inference · unmodified model weights.**

<a href="https://github.com/cv/laya-plays-smb3/releases/download/v0.1.0-world1-clear/laya-world1-1.mp4"><img src="docs/assets/level-clear.gif" alt="Recorded ending of the verified World 1-1 run: Mario jumps into the goal card and the game displays COURSE CLEAR" width="512"></a>

*Real emulator footage at original game speed. Text observations from visible-state RAM;
fixed rightward movement, model-driven jumping. One selected stochastic run in stepped mode.*

[Watch the full run](https://github.com/cv/laya-plays-smb3/releases/download/v0.1.0-world1-clear/laya-world1-1.mp4) · [Release & evidence](https://github.com/cv/laya-plays-smb3/releases/tag/v0.1.0-world1-clear) · [Reproduce it](#quick-start) · [How it works](#how-it-works)

</div>

## Why put a decision model in Mario?

Because a probability distribution is much more interesting when the wrong answer walks into a pit.

This is a small experimental playground for **model-driven game control**: native Laya inference
on an NVIDIA DGX Spark, real NES emulation, structured observations, and an unusually inspectable
trail of decisions. The interesting part isn't just whether Mario survives—it's seeing **what the
model was told, which action it favored, and whether that action arrived in time**.

- **Real SMB3, not a look-alike.** Stable Retro + FCEUmm run the game; the harness owns the control