Tag: chess engines

The history of Stockfish: from kitchen-sink code to chess powerhouse

Stockfish is one of the great success stories of open-source software: a community-built chess engine that began as a hobby fork and grew to dominate engine competitions, power analysis on the biggest chess sites, and push the boundaries of how engines are built. Here’s a concise history: how it began, how it improved, what it is today, and whether the “fish” has ever actually been beaten.

How it began

Stockfish’s roots go back to Glaurung, an engine written by Norwegian programmer Tord Romstad (released 2004). In 2008 Italian programmer Marco Costalba forked Glaurung, and the new project was named Stockfish — “produced in Norway and cooked in Italy,” as the joking origin line goes. The first Stockfish release (1.0) landed in November 2008; over the next few years Stockfish and Glaurung shared ideas until Stockfish became the primary development track. From the start it was open-source and community-driven, which set the tone for how it would scale.

How it improved (tech and community)

Two forces explain Stockfish’s rise: relentless engineering and a collaborative community.

1. Classic engine improvements — Stockfish’s core relied on highly optimized alpha–beta search, handcrafted evaluation features (king safety, pawn structure, piece-square tables), and vast tuning against test suites and self-play. Contributors constantly optimized low-level code, parallel search, move ordering, and tablebase usage.
2. Adoption of neural techniques (NNUE) — Historically Stockfish was a “classical” engine (handcrafted eval + brute-force search). In 2020 Stockfish integrated NNUE (efficiently updatable neural-network evaluation), a technique imported from shogi research and adapted to chess. NNUE lets the engine combine fast neural-network positional evaluation with Stockfish’s search, giving a large Elo leap without sacrificing speed. The NNUE integration and subsequent hybrid/evolution of evaluations produced one of the biggest single gains in Stockfish’s modern history. Over time the project moved from a primarily classical evaluator to a neural-hybrid default.
3. Massive CI and testing culture — Because the project is open-source, proposed changes are run against huge test suites and thousands of self-play games. This continuous testing (plus rapid iteration from many contributors) means improvements are validated before becoming defaults.

How it is today: competitions, rankings, and the web

Stockfish is widely regarded as one of the strongest—or the strongest—open-source engines in the world. It has dominated the major engine contest scene (Top Chess Engine Championship — TCEC) repeatedly and has been a frequent winner of Chess.com’s engine events. These competition victories are why many consider Stockfish the default “world-class” engine in practice: it faces the very best engines in sustained automated tournaments and usually comes out on top.

Beyond tournaments, Stockfish is everywhere in everyday chess:

• Lichess uses Stockfish to provide game analysis and “computer analysis” tools for players (Stockfish runs in the cloud and in-browser variants are used for local analysis tooling). Lichess publicly discusses its use of Stockfish for analysis and studies.
• Chess.com uses Stockfish in various ways: Stockfish powers much of the analysis/back-end tools and Stockfish-based assessments are a major part of game review; Chess.com also runs its own engine tournaments in which Stockfish has historically dominated. (Chess.com’s infrastructure mixes engines and internal systems, but Stockfish is an important analysis component.)

In short: Stockfish is both a competition champion and the analysis engine millions of players see when they click “analyze my game.”

Has Stockfish ever lost?

Stockfish loses games like any top engine: in head-to-head matches with other super-strong engines, in specially configured matches, and historically in encounters with new paradigms of AI.

AlphaZero (2017)

One of the most famous moments was AlphaZero vs Stockfish (2017). Google DeepMind’s AlphaZero — a reinforcement-learning, self-learning neural net — played a published match against Stockfish that drew huge attention because AlphaZero won convincingly and exhibited highly unorthodox but deeply effective play. Those games became a cultural touchstone: not necessarily a definitive “Stockfish is weak” verdict, but a demonstration that self-trained neural approaches could produce a dramatically different and very strong style. (Readers can find annotated games from that match online.)

Leela Chess Zero and other neural engines

Following AlphaZero, Leela Chess Zero (LCZero) — an open-source project inspired by AlphaZero — emerged and began to rival Stockfish in TCEC and other events. There have been notable matches and upsets where LCZero beat Stockfish in long tournament runs; conversely, Stockfish has repeatedly adapted and reclaimed titles. Engine competitions since 2018–2020 have been a dynamic back-and-forth between tuned classical/hybrid engines (Stockfish) and neural net engines (LCZero and others). Top Chess Engine Championship seasons show this rivalry in the winners’ list.

Has it ever lost to a human?

So far, in fully fair conditions (classical or standard time controls, full hardware, no limitations), there is no well-documented case of a human beating modern Stockfish at full strength. Because its strength eclipses even the very best human players under normal settings, humans generally lose or draw engines.

In terms of raw strength, the rating gap makes clear why humans haven’t beaten Stockfish under full conditions. Magnus Carlsen’s Elo currently peaks around 2850 Elo on the FIDE rating list. In contrast, Stockfish’s engine Elo on rating lists like CCRL or CEGT is often 3700–3900 Elo (sometimes estimated over 4000 with optimal hardware). That’s a gulf of nearly 1,000 Elo points, far larger than the gap between Carlsen and an average grandmaster. In practical terms, it means even the best human might only manage an occasional draw against Stockfish, while consistent victories are currently virtually impossible without engine limitations.