Stockfish is a free and open-source chess engine, renowned for its strength and efficiency. Developed primarily in C++ and released under the GNU General Public License, it consistently ranks among the world’s strongest chess engines, competing in top computer chess tournaments. Stockfish employs a sophisticated alpha-beta search algorithm with a highly tuned neural-network-based evaluation (NNUE) since version 12, allowing it to run effectively on consumer hardware. Its development is community-driven, with contributions from volunteers worldwide, and it serves as both a competitive engine and a research platform for artificial intelligence and game theory.
1 History
1.1 Origins and early development (2008–2010)
Stockfish was created in 2008 by Tord Romstad, Marco Costalba, and Joona Kiiski, building upon the open-source GluChess engine. The name "Stockfish" refers to a dried fish that keeps for a long time, reflecting the developers’ intention to create a durable and continually improving project. The engine was initially released under the GNU GPL and quickly gained a following due to its strong play and transparent development process. Early versions relied on a handcrafted evaluation function and a standard alpha-beta search, achieving competitive strength among top open-source engines.
1.2 Rise to top strength (2011–2015)
By 2011, Stockfish had surpassed many proprietary engines, winning its first major tournament title at the 2011 Top Chess Engine Championship (TCEC). The project’s rapid improvement was driven by a large community of contributors submitting patches through a rigorous testing framework. Key advances during this period included better move ordering, improved pruning techniques, and more efficient search algorithms. Stockfish’s rating on the Computer Chess Rating List (CCRL) steadily climbed, placing it among the top three engines in the world by 2014.
1.3 Introduction of NNUE (2020–present)
In 2020, Stockfish integrated a neural-network evaluation function (NNUE, Efficiently Updatable Neural Network), a technology pioneered by the engine Shogi (Japanese chess) and later adopted by the chess engine NNUE. The NNUE evaluation dramatically improved positional understanding and tactical accuracy. Stockfish version 12, released in December 2020, marked the first official release with NNUE support. Subsequent versions (13 and 14) continued to refine the network architecture and training methodology, maintaining Stockfish’s position as the strongest chess engine on consumer hardware. The NNUE component allowed Stockfish to maintain high performance on CPUs without specialized hardware like GPUs.
2 Technical architecture
2.1 Search algorithm
2.1.1 Alpha-beta pruning
Stockfish uses the alpha-beta pruning algorithm to reduce the number of nodes explored in the game tree. This algorithm eliminates branches that cannot influence the final decision, allowing deeper searches within the same time budget. Stockfish implements several enhancements, including iterative deepening, aspiration windows, and principal variation search, to maximize pruning efficiency.
2.1.2 Null-move pruning and reductions
Null-move pruning allows Stockfish to skip a side’s move and see if the opponent can make significant progress; if not, the position is considered safe and the search depth is reduced. This technique saves computational effort on dead or boring positions. Stockfish also employs futility pruning, late move reductions (LMR), and other reduction heuristics to further trim the search tree without sacrificing accuracy.
2.1.3 Transposition table and hash management
A transposition table (TT) stores previously evaluated positions along with their scores and best moves. Stockfish uses a large hash table (configurable by the user) to avoid re-searching identical positions that arise via different move orders. The TT is managed with aging and replacement strategies to preserve relevant entries and handle memory constraints. Hash collisions are detected using checksums and the table is cleared for new games to prevent contamination.
2.2 Evaluation function
2.2.1 Handcrafted evaluation (pre-NNUE)
Before version 12, Stockfish relied on a manually tuned evaluation function composed of material counts, piece-square tables, king safety, pawn structure, mobility, and other chess knowledge. The evaluation weights were determined by human experts and adjusted based on statistical analysis of self-play games. This system was highly effective but eventually reached a plateau where further improvements became difficult.
2.2.2 Neural-network evaluation (NNUE)
2.2.2.1 Network structure and inputs
The NNUE evaluation function is a fully connected neural network with a single hidden layer of 256 neurons. Input to the network consists of features representing piece positions, king safety, and other board characteristics, encoded in a sparse representation. The network uses halfKP (king + piece) input features, meaning each feature is defined by the combination of a piece’s type, color, and relative position to the two kings. This sparsity allows efficient incremental updates during search (hence "efficiently updatable").
2.2.2.2 Training data and methodology
The neural network is trained on millions of positions from high-quality self-play games generated by Stockfish itself. Training uses supervised learning with gradient descent, where the target is the score from a deep search (e.g., a full search of depth 20+). The training procedure periodically retrains the network on fresh data to adapt to evolving search improvements. The network weights are quantized to 16-bit integers for fast inference on CPU.
2.3 Multithreading and scalability
2.3.1 Split-point and lazy SMP
Stockfish supports parallel search via two main approaches: split-point (splitting the search tree across threads) and lazy SMP (Simultaneous Multi-Threading). In split-point, when a thread reaches a certain depth, it may split the remaining search among idle threads. Lazy SMP, adopted more recently, allows multiple threads to run independent searches with only occasional synchronization via shared transposition tables. This reduces overhead and scales well on many-core machines.
2.3.2 NUMA-aware optimizations
To improve performance on multi-socket systems (Non-Uniform Memory Access, NUMA), Stockfish includes NUMA-aware memory allocation and thread affinity settings. The engine can bind threads to specific memory nodes and cores to minimize memory access latency. These optimizations are enabled automatically when Stockfish detects a NUMA architecture.
3 Features and variants
3.1 Universal Chess Interface (UCI) protocol
Stockfish implements the Universal Chess Interface (UCI) protocol, enabling it to communicate with a wide variety of graphical user interfaces (GUIs) and chess platforms. UCI commands allow users to set search parameters, adjust time controls, configure engine options, and query analysis results. Stockfish supports all standard UCI commands as well as several proprietary extensions for greater flexibility.
3.2 Supported chess variants
3.2.1 Fischer random chess (Chess960)
Stockfish fully supports Fischer Random Chess (Chess960), where the starting position is randomized to reduce reliance on opening theory. The engine handles the non-standard castling rules and piece placements without performance penalty. Users can enable Chess960 mode via UCI settings.
3.2.2 Atomic, crazyhouse, and other Fairy-Stockfish variants
Through the Fairy-Stockfish fork, Stockfish supports numerous fairy chess variants, including Atomic (where captures destroy surrounding pieces), Crazyhouse (where captured pieces can be dropped), and many more. These variants are not part of the main Stockfish project but are maintained by a separate developer team.
3.3 Analysis and tuning tools
3.3.1 Contempt and draw value settings
3.3.1.1 UCI_Contempt parameter
The UCI_Contempt parameter adjusts the engine’s preference for winning or drawing. A positive contempt value makes the engine avoid draws, while a negative value encourages drawing. This is useful for tournament play where aggressive or conservative styles are desired. The parameter is automatically reduced in certain endgame situations to prevent unrealistic play.
3.4 Analysis and tuning tools (continued)
Stockfish provides several built-in analysis tools, including perft (performance test) for move generation speed measurement, bench for benchmarking, and eval for evaluating a given position. It also supports the "UCI_AnalyseMode" command for prolonged analysis, and "UCI_WDL" for win-draw-loss probability estimation.
4 Usage and ecosystem
4.1 Integration with graphical interfaces
4.1.1 Arena, Cute Chess, and Lichess
Stockfish can be used with popular chess GUIs such as Arena (Windows), Cute Chess (cross-platform), and the Lichess online platform (via its built-in engine). Users simply load the Stockfish binary into the GUI’s engine list. On Lichess, Stockfish is the default analysis engine and is available for cloud analysis of user games.
4.2 Stockfish as a training partner
4.2.1 Playing strength levels (UCI_LimitStrength)
Stockfish can simulate weaker opponents by limiting its search depth or using the UCI_LimitStrength command. This allows human players to set the engine’s strength from beginner to grandmaster level. The engine scales its playing ability smoothly, making it a versatile training partner.
4.3 Stockfish in online platforms
4.3.1 Lichess cloud analysis and API
The Lichess platform uses Stockfish for cloud-based analysis of games played on the site. Users can request engine evaluation of positions, move suggestions, and full game reports. Additionally, Lichess provides a public API that allows third-party applications to access Stockfish analysis remotely, subject to usage limits.
5 Community and governance
5.1 Open-source development model
5.1.1 GitHub repository and pull request workflow
Stockfish is developed on GitHub, where anyone can submit patches via pull requests. Each change is reviewed by maintainers and must pass automated tests before being merged. The repository includes comprehensive documentation for contributors, including coding style guidelines and a testing checklist.
5.1.2 Testing framework (Fishtest)
Fishtest is Stockfish’s distributed testing framework, used to evaluate proposed changes against the current development version. Volunteers run thousands of games on their own computers to statistically determine whether a patch improves strength. Fishtest uses Elo rating calculations and a confidence threshold to accept or reject patches, ensuring that only proven improvements are integrated.
5.2 Governance and steering committee
5.2.1 Stockfish project maintainers
The Stockfish project is guided by a steering committee composed of long-term contributors. The committee decides on roadmap priorities, major technical decisions, and release schedules. Maintainers are elected by the community and serve indefinite terms; they can resign or be replaced by consensus.
5.3 Collaboration with hardware vendors
5.3.1 Intel, AMD, and ARM optimizations
Stockfish engages with hardware vendors to optimize performance on their platforms. Vendor engineers have contributed code for utilizing new instruction sets (e.g., AVX-512, SSE4.2) and for improving cache behavior. On ARM architectures (e.g., Apple M1), Stockfish includes specific tuning for efficiency and multithreading. These collaborations benefit both the engine and the vendors by showcasing hardware capabilities.
6 Performance and recognition
6.1 Top Chess Engine Championship (TCEC) results
Stockfish has been a dominant force in the Top Chess Engine Championship (TCEC), winning multiple season titles (e.g., Season 6, 9, 10, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24). It has consistently reached the finals, often defeating competitors such as Leela Chess Zero and Komodo. Its TCEC performance has established it as the strongest traditional chess engine in tournament conditions.
6.2 Computer Chess Rating List (CCRL) and other rankings
On the Computer Chess Rating List (CCRL), Stockfish regularly holds the top position among engines that run on standard consumer hardware. It also ranks high on the IPON and Fast Ratings lists. As of 2024, its estimated Elo rating exceeds 3600 on the CCRL 40/4 time control.
6.3 Benchmarks and Elo progression
6.3.1 Relative strength over versions
Stockfish has exhibited steady Elo gains with each major version. From version 10 to 14, the engine gained roughly 200–300 Elo points, with the NNUE introduction in version 12 providing the largest single leap (about 100 points). Benchmarks run on standard hardware (e.g., 4-core CPU) show node-per-second improvements of 10–20% per year due to algorithmic and micro-optimization enhancements.
7 Relationship with other engines
7.1 Derivatives and forks
7.1.1 Komodo, Houdini, and open-source variants
Stockfish has inspired several forks and derivatives. Komodo and Houdini are commercial engines that originated from earlier versions of Stockfish but later diverged with proprietary improvements. Open-source variants such as Fairy-Stockfish (supporting chess variants) and Brainfish (using Stockfish’s search with Leela’s evaluation) also rely on Stockfish’s core code. The project’s permissive license (GPL) allows these derivatives while requiring any modifications to be open-sourced when distributed.
7.2 Influence on AI research
7.2.1 Use in reinforcement learning studies
Stockfish serves as a baseline and opponent in reinforcement learning research for chess. Its fast, deterministic play makes it ideal for generating training data and evaluating learned agents. Studies such as those exploring neural-guided search or self-play reinforcement learning often compare against Stockfish. The engine’s open-source nature facilitates integration and reproducibility.
8 Licensing and distribution
8.1 GNU General Public License (GPL v3)
Stockfish is distributed under the GNU General Public License version 3 (GPLv3). This license ensures that users have the freedom to run, study, share, and modify the software. Any derivative work distributed publicly must also be open-sourced under GPLv3. The license does not impose restrictions on non-commercial use or personal modifications.
8.2 Binary releases and platforms
8.2.1 Windows, Linux, macOS, Android, and iOS
Official Stockfish binaries are provided for Windows (64-bit and 32-bit), Linux (64-bit and ARM), and macOS (Intel and Apple Silicon). Community-maintained ports are available for Android (via the Stockfish app on Google Play) and iOS (via the Stockfish app on the App Store). The engine can also be compiled from source for other operating systems.
8.3 Privacy and telemetry considerations
Stockfish does not collect or transmit any personal data. It operates entirely offline, with no telemetry, analytics, or network activity unless explicitly enabled by the user (e.g., via Lichess cloud analysis). The engine’s source code is auditable, ensuring transparency. Users concerned about privacy can run the engine in an isolated environment or check the code themselves.