Stockfish analysis, in your browser

Stockfish reports a score in pawns from White's point of view. +1.3 means White is better by roughly a pawn and a third of positional value; -0.4 means Black is a little more comfortable. When it finds a forced mate it shows M5 or -M3 instead: the number of moves until checkmate with best play.

Each line is a principal variation: the sequence the engine expects both sides to play. The first line is the engine's recommendation. The second and third lines show alternatives, and the gap between them tells you how critical the position is. If line one is +0.2 and line two is -2.1, there is exactly one good move.

Depth is how many half-moves ahead the main search has gone. Because Stockfish extends forcing sequences and prunes hopeless ones, a depth-20 search often sees tactics thirty or forty plies deep, while brushing past quiet moves it has judged irrelevant.

Reading the arrows

Green is the top choice, blue the second, yellow the third. Click any first move in the line list to play it on the board; the engine then analyses the new position immediately. Undo takes you back. This is the fastest way to check an idea: play your candidate move and see whether the evaluation holds.

Until 2020 Stockfish evaluated positions with hand-written rules: material, pawn structure, king safety, mobility, each with coefficients tuned over years. In August 2020 the team merged an “efficiently updatable neural network” (NNUE), originally developed for computer shogi, which learned its own evaluation from hundreds of millions of positions. The jump in strength was about 80 Elo overnight, larger than any single change in the engine's history.

The network is small by machine-learning standards, which is exactly why it works in a browser. The build here uses the compact network that the Stockfish project publishes for low-memory environments. It is a few dozen Elo weaker than the full network at fixed depth, a difference that matters in engine-versus-engine matches and not at all when the question is whether your knight sacrifice on move 23 was sound.

Limits worth knowing

• No tablebases. Positions with seven or fewer pieces are solved by search, not looked up, so the engine may show +9 where a tablebase says mate in 40.
• Mobile phones run the engine at a fraction of desktop speed. Depth 14 on a phone takes about as long as depth 18 on a laptop.
• The engine judges positions, not plans. A +0.3 line can be a nightmare to play for a human. Use it to check moves, not to choose a style.