Mahjonglm-10m: The AI That Plays Mahjong (and It's Actually Kind of Cool)

First, the obvious part: it's tiny. Ten million parameters. To put that in perspective, the big language models like GPT have billions of them. Translation: mahjonglm was designed to run on regular hardware—you don't need a thousand-dollar GPU gathering dust on your desk. It's the kind of model you download and test on your laptop without your CPU fans staging a protest.

But why mahjong specifically? Because it's a beast. You need memory, tactics, pattern recognition, and a dash of psychological bluffing. It's nothing like chess, where the rules are clean and predictable. Mahjong is a shape-shifter—combinations flow into each other, rules change depending on where you're playing, and you're constantly guessing what's hiding in your opponents' hands. It's a solid stress-test for seeing how well an AI can reason through messy, unstructured situations.

The model is purpose-built for mahjong (the "ccc" in the name probably refers to specific tile configurations), which means you're not using this to draft an email or summarize your favorite article. It's a specialist tool. Looking for a do-everything model? Keep scrolling. Want an AI that actually gets the game? This is your shot.

Here's the cherry on top: it's open source on HuggingFace. Anyone can grab the code, the architecture, the weights—everything. No paywalls, no proprietary API gatekeeping. If you're a developer with an itch to teach an AI how to navigate genuinely complicated games, you've got the whole toolkit.

With ten million parameters, don't expect world-championship performance. It probably plays reasonably—better than a beginner, worse than someone who actually knows what they're doing. It's a sandbox for learning and experimenting, not a threat to professional mahjong players. (And honestly, strong mahjong AIs already exist—they're much bigger and nobody's losing sleep over them.)