2048 AI

How to play 2048

2048 is played on a 4 x 4 board. Swipe (or press an arrow key) and every tile slides as far as it can in that direction. When two tiles with the same number meet, they merge into one tile with their sum: two 2s make a 4, two 4s make an 8, and so on. After every move that changes the board, a new tile appears on a random empty square: a 2 nine times out of ten, a 4 the rest of the time. Your score goes up by the value of every merged tile. Make a 2048 tile to win, then keep going if you like. The game ends when the board is full and no move can merge anything.

Two rules catch people out. A tile merges only once per move, so a row of 2, 2, 2, 2 moved left becomes 4, 4, not 8. And when three equal tiles line up, the two nearest the wall you are moving toward merge first: 2, 2, 2 moved left gives 4, 2. You can undo one move at a time, and the game is saved in your browser, so a reload does not lose it.

The AI helper

Suggest move shows the move the AI would make, with how the other moves compare. AI plays lets it play on its own; pick the speed, and press any arrow key or Stop AI to take over. AI strength sets how far ahead it looks: Quick searches up to three moves ahead, Strong up to four. Strong plays better and takes longer per move, which you only notice at Max speed or on an older phone.

Here is how it did in our tests, from fixed starting positions, with no undo:

So it does not win every game, and nobody can promise that: the new tiles land at random, and now and then they land badly.

How the 2048 AI works

The AI uses expectimax search. For each of the four moves it asks what could happen next: the new tile could land on any empty square and be a 2 or a 4, so it averages over all of those, with the real odds of 9 to 1. Then it tries its best reply to each of those boards, and so on, a few moves deep. Branches that are very unlikely (under 1 in 10,000) are cut short to save time. At the end of each line it scores the board with a well-known heuristic from an open-source 2048 AI by nneonneo: points for empty squares and for equal tiles side by side, a penalty for rows and columns that do not rise or fall steadily, and a penalty for large tiles in general. The move with the best average is the one it plays. It knows only what you know: it cannot see where the next tile will appear.

2048 solver: get the best move for your own board

Playing on another site or app? Switch to Solver, pick a number from the palette and tap the squares to copy your board, then press Best move. You get the move to make and how the others compare. Play from this board continues the game here, and Copy current game brings your game board into the editor.

Strategy tips

Pick a corner for your biggest tile and keep it there. Keep the row along that wall full so the tile cannot be pushed out, and build the numbers in that row in order, biggest at the corner, so merges run toward it. Use mostly two directions, a third when you must, and the fourth only as a last resort. Merge small tiles early before they scatter, and look at what a move does to the empty squares: a board with room to breathe is a board that survives. Watching the AI play at Slow speed is a good way to see these habits in action.

Like 2048, the block puzzle solver is about keeping the board open: it finds the best order and spot for your three pieces.

Solver tested 2 October 2026. Found a problem? Tell us.

Questions