How a hobby fork became the strongest chess player on Earth
Stockfish has no company behind it, no marketing budget, and no secrets. It is a GPL-licensed program maintained by volunteers, and it has been the reference point for computer chess for more than a decade. The build running this exchange is compiled straight from the official repository: loading….
- 2004
Glaurung
Norwegian programmer Tord Romstad releases Glaurung, an open-source engine named after Tolkien's dragon. It is respectable, not dominant, and it is free for anyone to study and modify. That decision is the seed of everything that follows.
- 2008
The fork
Italian developer Marco Costalba forks Glaurung 2.1 and releases Stockfish 1.0. The name is a wink at the two founders' countries: stockfish is Norwegian dried cod, a staple of Italian cooking for centuries. Romstad and Joona Kiiski join, and the three become the project's founders.
- 2013
Fishtest
Gary Linscott builds Fishtest, a distributed testing framework. Volunteers donate CPU time, and every proposed change must win a statistically rigorous test over tens of thousands of games before it is merged. Intuition stops mattering. Only measured strength counts.
- 2014
First TCEC title
Stockfish wins Season 6 of the Top Chess Engine Championship, the de facto world championship for engines. It has been in nearly every superfinal since, and has won the majority of them.
- 2017
AlphaZero
DeepMind's AlphaZero, trained purely by self-play with a neural network, beats Stockfish 8 in a widely publicised match. The conditions are disputed, but the message lands: neural evaluation is the future. The open-source community answers with Leela Chess Zero, and the two projects push each other for years.
- 2020
NNUE
Stockfish 12 ships with NNUE, the Efficiently Updatable Neural Network invented by Yu Nasu for shogi and ported to chess by Hisayori Noda. It runs on an ordinary CPU, updates incrementally as pieces move, and adds roughly a hundred Elo overnight.
- 2021–22
Defending the licence
The project sues ChessBase over GPL violations in a commercial engine built on Stockfish code. The case is settled in 2022. A rare moment of an all-volunteer project enforcing the rules that made it possible.
- 2023–24
All neural
The last remnants of the classical, hand-written evaluation are removed. Every position is now judged by a network trained on billions of positions, searched by an alpha-beta engine tuned by Fishtest to within a few Elo of perfection.
- Today
Still open, still winning
Hundreds of contributors, thousands of Fishtest workers, a rating far beyond any human's. And on this exchange, handicapped to a level you can actually beat.
Search
Alpha-beta search with decades of pruning tricks, examining millions of positions per second on a laptop. It routinely looks 30 to 40 moves deep on the lines that matter.
NNUE evaluation
A small neural network that scores positions in microseconds. Because pieces move one at a time, the first layer is updated incrementally instead of recomputed, which is what makes it fast enough for CPU search.
Fishtest
Nothing is merged on opinion. Every patch plays tens of thousands of games on volunteer machines and must pass a sequential probability ratio test.
Levels 0–5
Blunders like a beginner, hangs pieces, walks into mates. Pays 1.1× to 1.5×.
Levels 6–14
Club strength and up. It punishes loose play but still makes real mistakes. Pays 1.7× to 6×.
Levels 15–20
Expert to superhuman. Wins are rare enough that the house pays 8× to 30×. Level 20 is the unhandicapped engine.