Kimi 3 Took 2 Hours to Build Pac-Man—Then Pac-Man Got Stuck

After lavish praise, enormous expectations and breathless claims about Kimi 3’s coding abilities, I decided to subscribe and put the newly released AI model to the test myself.

I was not asking it to build the next Grand Theft Auto. I did not request a massive open-world adventure, a complicated multiplayer server or a revolutionary new game engine.

I asked it to make Pac-Man.

My exact prompt was:

“Create a Pac-Man game clone using a single HTML file. Let’s make the visuals, SFX and gameplay identical.”

Simple enough—or so I thought.

THE TWO-HOUR PAC-MAN MARATHON BEGINS

Kimi 3 got to work generating the game, writing code and repeatedly attempting to piece together its single-file arcade masterpiece.

Then I waited.

And waited.

More than two hours later, the much-hyped AI coding machine finally delivered its Pac-Man-style creation.

To be fair, the game loaded. It had a maze, familiar-looking enemies, sound effects and a yellow character moving around the screen eating dots.

For a brief moment, it appeared Kimi 3 had pulled it off.

THEN PAC-MAN GOT STUCK

After actually playing the game, the problems quickly started showing themselves.

The main character became stuck inside the maze, unable to move correctly through parts of the level. The gameplay that was supposed to be “identical” to the arcade classic clearly was not.

After more than two hours of AI-powered coding, testing and generation, Kimi 3 had produced a Pac-Man clone in which Pac-Man apparently could not reliably navigate his own maze.

That is a fairly serious problem when navigating the maze is essentially the entire game.

THE HYPE TRAIN HITS A WALL

Kimi 3 arrived surrounded by glowing praise, impressive benchmark claims and predictions that it could become one of the most powerful coding models available.

That made my first real test especially disappointing.

A single-file Pac-Man clone should be an ideal task for a supposedly advanced coding AI. The rules are established, the mechanics are simple and countless examples of similar browser games already exist.

Instead, the process stretched beyond two hours and ended with a game-breaking movement issue.

Could a human developer fix it? Almost certainly.

Could Kimi 3 eventually repair the game after several more prompts? Probably.

But that is not quite the effortless, autonomous coding revolution people may expect after reading all the glowing coverage.

NICE TRY, KIMI

This does not mean Kimi 3 is useless. New AI models often improve rapidly, and one coding test cannot measure every capability a system has.

However, first impressions matter—especially when users are paying for access after hearing extravagant praise about a model’s intelligence and coding power.

My first impression was watching an advanced AI spend more than two hours making a basic arcade clone, only for Pac-Man to get trapped in the maze.

And it used over 10% of my sub plan :))

Nice try, Kimi.

But perhaps do not give up the day job just yet.