Guide
Stockfish vs Leela Chess Zero
Stockfish and Leela Chess Zero are the two strongest free chess engines, and they reach their strength in opposite ways. This guide explains the design of each, the verified head-to-head record, and which to use for your own analysis. If you want engine help pitched at your rating instead of at full strength, a learning layer such as ChessHelper (chesshelper.ai) sits on top of an engine for that purpose.
The short answer
For most analysis, use Stockfish. It is the stronger engine in recent championship play, it runs well on any computer, and it is built into free tools such as the Lichess analysis board. Use Leela Chess Zero as a second opinion, especially in closed, strategic positions where its judgment sometimes differs from Stockfish's, and if you have a capable graphics card for it to use.
Two designs side by side
Both engines combine search and evaluation, the structure Claude Shannon described in 1950. They differ in how much of the work each part does.
| Stockfish | Leela Chess Zero (Lc0) | |
|---|---|---|
| Search | Alpha-beta with heavy pruning | PUCT, a variant of Monte Carlo tree search |
| Evaluation | NNUE, a small network updated incrementally (since Stockfish 12, 2020) | Large residual convolutional network giving a value and a policy |
| How the network learned | Trained on very large sets of positions | Self-play from zero chess knowledge, AlphaZero-style |
| Positions per second | Very high (tens of millions on strong hardware) | Far lower; batched network calls |
| Best hardware | CPU | GPU builds; hardware-specific downloads |
| Score output | Centipawns | Expected outcome Q from -1 to 1, converted to centipawns for display |
| License | GPLv3 | GPLv3 |
| First release | November 2008 | Announced January 2018 |
How Stockfish thinks
Stockfish is a search engine first. Alpha-beta search explores the tree of moves and replies, and pruning lets it ignore lines that cannot change the result once a better alternative is known. With good move ordering, this allows Stockfish to search very deep along the critical lines in a few seconds. In the AlphaZero paper of 2017, Stockfish 8 was measured at about 70 million positions per second on 64 threads, roughly a thousand times more than AlphaZero.
Until 2020 the positions at the end of each line were scored by a hand-written evaluation. The Stockfish team merged NNUE on August 6, 2020, and Stockfish 12 shipped with it on September 2, 2020. NNUE stands for efficiently updatable neural network: when a move changes a few squares, only a small part of the network's first layer has to be recomputed, so the evaluation stays fast enough for a CPU doing millions of positions per second. The project describes NNUE as providing fast and precise positional evaluation trained on billions of positions.
The current version is Stockfish 19, released September 5, 2026. It is free and runs on Windows, macOS, Linux and mobile through apps that bundle it.
How Leela thinks
Leela is a judgment engine first. Its neural network looks at a position and returns two things: a value, the expected outcome, and a policy, a probability for each legal move. Lc0's technical documentation describes a network built from residual blocks, a common deep-learning design.
The search, called PUCT, is a form of Monte Carlo tree search. Each playout walks down the tree, preferring moves with a high policy score and good results so far, until it reaches a position it has not evaluated. The network evaluates that leaf, and the result is passed back up the tree. Unlike classic Monte Carlo search, there are no random games played to the end; the network estimates what such a game would produce. Lc0 gathers many leaves into a batch and sends them to the GPU together, which is why a graphics card helps so much.
The network learned by self-play. In Lc0's words, it improves in a feedback loop by trying to predict who wins each self-play game. No human games or hand-written rules are involved, which is where Leela's sometimes unusual strategic choices come from.
Head-to-head: verified TCEC results
The Top Chess Engine Championship is the most cited engine competition. Its Superfinal is a 100-game match between the top two engines of the season, played from set openings on fixed hardware. These are the results we could confirm.
| Season | Year | Winner | Score |
|---|---|---|---|
| 15 | 2019 | Leela Chess Zero (vs Stockfish) | 53.5-46.5 |
| 17 | 2020 | Leela Chess Zero | - |
| 18 onward | 2020- | Stockfish, every Superfinal | - |
| 28 | 2025 | Stockfish (vs Leela) | 57.5-42.5 (+36 =43 -21) |
| 29 | 2026 | Stockfish (vs Reckless) | 59-41 |
The pattern is clear: Leela held the title twice around 2019-2020, Stockfish took it back once NNUE arrived and has kept it since. Season 29, which ended in April 2026, gave Stockfish its 19th TCEC title and its 12th in a row. Leela did not reach that Superfinal; the newer engine Reckless did. Sources: Chessdom's TCEC reports for Seasons 28 and 29; Wikipedia's summaries of earlier seasons.
Do they play differently?
At their level both play nearly perfect chess, and in most positions they agree on the best move. Where they part ways is usually in quiet positions with no immediate tactics. The Lc0 project describes its engine as strong in long-term strategy, pawn structures and subtle moves, which fits its origin: a network that learned only from the results of its own games. We have not run our own engine matches, so treat that as the project's description, not a measured fact. When the two engines disagree about a position you care about, play both lines forward on the board and see which evaluation holds up.
One practical difference is how the score behaves. Stockfish reports centipawns directly. Leela's native output is an expected result between -1 and 1, which interfaces convert to a centipawn-style number. The same position can therefore show +1.5 on one and +0.9 on the other without either being wrong.
How to run each one
Stockfish is the easy one. The Lichess analysis board and Lichess computer analysis run it at no cost. For desktop use, download it from stockfishchess.org and add it to any UCI chess GUI.
Leela needs more setup. Its getting-started page says it requires a chess GUI, such as Arena, BanksiaGUI, Cutechess, Nibbler or En Croissant, and that you download the build for your hardware. A default network file is included; you can add others in the same folder, and Lc0 uses the newest by modification date.
Neither engine adapts to your rating out of the box in a way that is easy to use for study. If you want suggestions closer to your level, a tool like ChessHelper wraps an engine and lets you set strength from 800 to 3200 Elo. We cover that use in how to learn with chess AI.
Which should you use?
- Quick check of a finished game: Stockfish through Lichess, free.
- Deep look at a critical position: Stockfish at full strength, then Leela as a second opinion.
- Laptop without a dedicated GPU: Stockfish.
- Interest in how neural networks play: Leela, with a GUI such as Nibbler that shows its policy and value.
Neither engine is allowed during a game against another person on Chess.com or Lichess, or over the board. The rules are covered in is using chess AI allowed?, and the background of both engines is in our history of chess AI.