What Is Stockfish? The Chess Engine Behind Modern Analysis, Explained

Stockfish is a free, open-source chess engine, and for most of the last decade it has been the strongest chess-playing program in the world. When a site shows you an evaluation bar, a best-move arrow, or an accuracy score, there is a good chance Stockfish is doing the work underneath.

It began in 2008, derived from an earlier engine called Glaurung, and grew into a large volunteer project. That openness is a big part of why it became so strong. Proposed changes are tested by a distributed system that plays them in huge numbers of games on computers donated by volunteers, and a change is only accepted if it measurably wins more. Improvement is empirical rather than a matter of opinion.

At its core Stockfish does two things. It searches — it looks ahead through the tree of possible moves, using aggressive pruning to skip lines that are obviously bad so it can go deeper on the ones that matter. And it evaluates — at the end of each line it looked at, it estimates who is better and by how much. The move it recommends is the one whose line leads to the best evaluation against the opponent's best replies.

The biggest change in its history was NNUE, a kind of neural network designed to be updated cheaply as pieces move, adopted in 2020. Before it, Stockfish's evaluation was a large hand-written formula of chess knowledge — material, king safety, pawn structure, mobility. The network learned to judge positions better than that formula, and it is fast enough to run on an ordinary CPU, so the engine got stronger without needing special hardware. Modern versions evaluate positions entirely with the network.

Its evaluations are shown in pawns. A score of +1.00 means White is better by roughly the value of a pawn; -1.00 means the same for Black. "M5" or "#5" means a forced checkmate in five moves. In practice, anything between about -0.3 and +0.3 is equal, and a club player should not read much into small differences inside that range.

Two things are easy to misread. The first is depth: an evaluation from a short search can change completely once the engine looks further, especially in tactical positions, so a number that appears instantly deserves less trust than one that has had time to settle. The second is practicality: Stockfish assumes perfect play from both sides. It will happily recommend a line that requires five only-moves in a row, and score your safer human choice as worse, even when the safer choice is what you should play over the board.

That is why an engine is best treated as a source of information rather than a judge. It is extraordinarily good at telling you what happens in a position. It cannot tell you why you chose the move you did, which idea you were missing, or whether the "best" move was one a person could realistically find in the time you had.

On ChessDeeper, Stockfish runs in your browser through WebAssembly rather than on a server, so reviewing a game never waits in a queue. It looks at two candidate moves in each position, which is how the review can tell a move that was simply good apart from one that was the only move that held the position together.