Every game here is played live by Gemma 4 E4B, a 4-billion-parameter open model running on the lab's RTX 5090, not a cloud API. Watch the panel on the right of each game: that is the model literally thinking out loud before every move. It generates at over a hundred tokens a second; we replay its reasoning at reading speed so you can follow along.
Turnip Hold'em
Heads-up mini poker against Gemma. She reads her hand out loud, makes honest bets, and yes, you can bluff her. Three hands, most chips wins.
Can you out-bluff it?Maze Race
Same maze, two runners. You see the whole map. Gemma is inside the maze and can only see the hedges next to her. Watch her reason her way out of dead ends.
Race the reasoningMind Reader
Think of anything. Gemma gets ten yes/no questions to figure it out, and shows you exactly how each answer narrows the field. Stump her and you win.
10 questions · she guessesWhat you're actually looking at
Small language models can do much more than autocomplete text. Given the rules and the game state as plain words, the same 4B model plans poker bets, navigates mazes and plays twenty questions. Because it's a reasoning model, we can stream its deliberation before each move. This is the kind of thing the club builds: our flagship workshop trains a model like this one to play the board game Codenames, from dataset to fine-tuning. Join us.