3D 2048 Rules Reference
The other guides on this site tell you what to do. This one tells you what the game does. Below are the merge, spawn, scoring, undo and game-over rules exactly as they are implemented, a specification table for the three modes, and the glossary the rest of the guides assume you already know.
This page contains no tactics on purpose. Corner pinning, chain shapes, Z-axis discipline, cascade rows and the endgame squeeze all live in the mode guides โ 2ร2ร2, 3ร3ร3 Normal, Hard. What you get here is the layer underneath them: the rules those tactics are built on, written down precisely so you can check a plan against them instead of guessing.
How a move actually resolves
Most players learn 2048 by feel and end up with a mental model that is roughly right and wrong in a few specific, expensive places. These are the places.
Three identical tiles in a line merge as two โ and you pick which two
Put three 4s in a straight line and slide along it. You do not get a 12, and you do not get two merges. You get an 8 and a leftover 4.
The reason is the order the line is resolved in. The game walks the line starting from the wall you are sliding toward and works backward. The first eligible pair it meets โ the two tiles closest to the destination wall โ merges. It then skips past the tile it just consumed, so the third tile cannot be pulled into the same merge. The leftover survives at full value, sitting immediately behind the new doubled tile.
The practical consequence is that the leftover is not chosen for you. With 4 ยท 4 ยท 4 in a line, sliding one way puts the 8 against that wall with the spare 4 behind it; sliding the other way puts the 8 against the opposite wall, with the spare 4 on the other side of it. Same two output tiles, mirrored positions. When you have three of a kind lined up, you are not deciding whether to merge โ you are deciding which end of the line the spare ends up on, and therefore what it is adjacent to next turn.
The same rule caps a single line at one merge per move, no matter how the values fall. There is no chain reaction inside one line.
A move that changes nothing costs nothing
The game performs the slide first and then asks whether any tile ended up in a different cell along that axis. If none did, the entire move is discarded: no tile spawns, the score does not move, and the undo slot is left alone. Pressing into a wall you are already flush against is free.
Be careful with the boundary, though. It is not "no merges happened" โ it is "nothing moved at all." If a single tile anywhere in the cube shifts by one cell, the move counts in full and the spawn fires.
Score counts merges, not moves
Every merge adds the value of the tile it produced. Two 2s make a 4 and add 4. Two 512s make a 1024 and add 1024. A move that rearranges half the cube without merging anything adds exactly zero.
So your score is a record of merge history, not of survival time, and two players sitting on the same highest tile can have very different scores depending on how much other material they merged on the way there. If you want the closed-form arithmetic behind that โ what a given top tile is "worth" and what the ceiling is โ that is the math page, not this one.
Undo is one level deep, and it re-rolls the spawn
Z undoes exactly one move. There is a single snapshot slot: making a move fills it, and undoing empties it. You can never undo twice in a row, and you cannot walk a run backwards.
The part that surprises people is what happens to the spawn. The snapshot is the board as it was before the slide, so undoing also removes the tile that spawned as part of that move. Replay the same move and the spawn is drawn fresh โ a new random empty cell, and a new random 2-or-4 roll.
Undo is therefore not only a correction tool. It is a re-roll. The same slide, repeated after an undo, can hand you a materially different board. What it will not change is the slide itself: which tiles move where and which pairs merge is fully determined by the position, so undoing and repeating an identical move only ever varies the spawn.
Best score never goes down
Undo restores the previous position's tiles, score and status, but the best score is deliberately kept at the higher of the two values, and that higher value is what gets saved. A move that set a new personal best followed immediately by an undo leaves the best untouched.
Which means there is no scoreboard cost to undoing a record-setting move. If the move that set your best also wrecked your position, you can take it back and keep the number.
Game over needs a full board and no adjacent pair
A full board is not a loss. The run ends only when both conditions hold at once: every cell is occupied, and there is no pair of equal values sitting next to each other along any of the three axes.
That distinction carries more weight than it sounds like. A completely packed 27-cell cube can stay playable for a long stretch, because one matching adjacent pair anywhere in it is enough to keep going. The late game is therefore not about protecting empty cells โ those are gone โ but about never letting your last legal pair get separated.
One ordering detail worth knowing: reaching the target tile is checked before the no-moves condition. If a single move both creates the target and leaves you with nothing legal, the run is recorded as a win. If you then choose to continue from a position that has no legal move, the game goes straight to game over rather than leaving you stuck pressing keys that do nothing.
The tile budget of a single move
Each merge turns two occupied cells into one, freeing exactly one cell. Then the spawn refills. That gives a fixed arithmetic per move:
| Merges in the move | Cells freed | Normal (1 spawn) | Hard (2 spawns) |
|---|---|---|---|
| 0 | 0 | +1 | +2 |
| 1 | 1 | 0 โ break even | +1 |
| 2 | 2 | โ1 | 0 โ break even |
| 3 | 3 | โ2 | โ1 |
Read the Normal column first, because it is the one nothing else on this site spells out: one merge per move exactly holds the board steady. Not "roughly" โ exactly. Every move you make with no merge has to be paid back later by a move with two, and there is no other currency. Hard shifts the whole column down by one, which is why it is a different game rather than a slower one; what to do about that is covered in the Hard mode guide.
Choosing a mode
Three modes, and the differences between them are entirely in this table. Nothing else about the rules changes.
| 2ร2ร2 | 3ร3ร3 Normal | 3ร3ร3 Hard | |
|---|---|---|---|
| Cells | 8 | 27 | 27 |
| Line length | 2 | 3 | 3 |
| Target tile | 256 | 2048 | 2048 |
| Tiles spawned per move | 1 | 1 | 2 |
| Chance a spawn is a 4 | 10% | 15% | 40% |
| Tiles on the opening board | 2 | 2 | 2 |
| Availability | โ | โ | 3ร3ร3 only |
Two details in that table are easy to miss. Hard mode's opening board is the same two tiles as Normal โ the doubled spawn only begins with your first move, so the first position you look at is not yet harder than usual. And Hard is not offered on the 2ร2ร2 board at all; selecting the small board silently puts you back on Normal.
Where a new player should start
Start on 2ร2ร2, for a reason that comes straight out of the budget table above rather than from taste. On eight cells with one spawn per move, the merge-or-drown arithmetic becomes visible within a handful of moves instead of a hundred. You find out whether your last five moves were affordable almost immediately, and a feedback loop that short is what actually teaches the rule. It also has the lowest chance of spawning a 4, so the board stays uniform longer and you can spend your attention on learning to see in three dimensions rather than on awkward leftover material.
What each board teaches that the previous one could not
On 2ร2ร2 every line is two cells long. Three identical tiles can never line up, so the first rule on this page โ the one about which two of three merge โ simply cannot occur. What you are learning there is more basic and more transferable: mapping what you see on screen to which cells a given key will actually move, reading a cube instead of a grid, and feeling the cost of a wasted move.
Move to 3ร3ร3 and lines become three cells long, so that rule switches on and starts applying on almost every turn. Along with it come the positional distinctions the 3ร3ร3 guide is built around, and a back layer large enough that you genuinely cannot hold it all in your head.
When to move to Hard
The honest signpost is qualitative: go to Hard once you can reach 2048 in Normal, not before. Hard is the same cube, the same target and the same rules โ the only thing it changes is the spawn column. So it has nothing new to teach you spatially. If Normal is still ending before 2048, Hard will end the same runs sooner and you will learn the same lesson with less time to observe it.
Switching costs you nothing you have already built, either. Each combination of board size and difficulty keeps its own saved game, its own best score and its own local ranking list. Trying Hard does not overwrite your Normal records, and an unfinished Normal run is still waiting when you switch back.
Glossary
Terms used consistently across every guide on this site.
- Slide direction. One of six: left / right, up / down, forward / backward. A slide always pushes every tile as far as it will go toward that wall. There is no partial move and no way to move a single tile.
- Merge. Two tiles of equal value, adjacent along the slide axis, combining into one tile of double value. Unequal values never merge, and a tile takes part in at most one merge per move.
- Spawn. The tile โ or in Hard, the two tiles โ added after a move that changed the board. The cell is chosen uniformly at random among empty cells; the value is 2 or 4 according to that mode's rate in the table above.
- Camera-relative controls. The four horizontal inputs are interpreted relative to where the camera is looking. The game takes the camera's facing direction, flattens it onto the horizontal plane and snaps it to the nearest cube axis, so "left" keeps meaning screen-left after you rotate. Up and down are the exception: they are not camera-relative and always map to the world vertical axis, because rotating the view never turns the cube upside down.
- Back-view window. The small panel that renders the same board a second time from a fixed viewpoint on the far side. It does not follow the main camera โ that is the entire point, since the two views between them cover the faces the other hides. You can drag it anywhere and resize it, and both stick between sessions.
- Vertex, edge, face-center, body-center. The position vocabulary for the 27-cell cube: 8 corner (vertex) cells, 12 edge cells, 6 face-center cells and 1 body-center cell. The 2ร2ร2 cube has only the first kind. Why the distinction matters tactically is covered in the 3ร3ร3 guide.
- Run. One game from the opening board to a win or a game over. Best score, saved game and local rankings are all stored per board size and difficulty, never pooled.
- Target tile and continue. The value that ends a run as a win โ 256 on 2ร2ร2, 2048 on 3ร3ร3. You may continue playing from there; the win is already recorded.
Which page answers which question
- 2ร2ร2 strategy โ what to do on the eight-cell board.
- 3ร3ร3 Normal strategy โ what to do on the twenty-seven-cell board.
- Hard mode strategy โ what to do when two tiles arrive every move.
- The math of 2048 โ the arithmetic: score formulas, probabilities, theoretical ceilings.
- FAQ โ controls, browser support, offline play, where your scores are kept.