Stockfish vs Leela is a contest between two designs. Stockfish runs a fast alpha-beta search on CPUs and judges positions with a small neural network (NNUE). Leela Chess Zero (Lc0) runs a much larger neural network on GPUs and searches far fewer positions, guided by the network's own move suggestions. As of October 2026 Stockfish is ahead: it beat Leela 57.5–42.5 in the TCEC Season 28 Superfinal, and on the CCRL Blitz list Stockfish 17.1 on a single CPU is rated 75 points above the best Lc0 entry. Leela has beaten Stockfish before, though, and the neural-network idea it borrowed from AlphaZero changed Stockfish itself.
Stockfish vs Lc0 at a glance
| Stockfish | Leela Chess Zero (Lc0) | |
|---|---|---|
| Search | Alpha-beta tree search with hand-designed heuristics | Tree search guided by the network's policy output |
| Evaluation | NNUE, added in Stockfish 12 (September 2020); hand-crafted eval removed July 2023 | Large network with policy and value outputs; transformer architecture since 2022 |
| Hardware | CPU | Designed for GPU; CPU backends since 2019 |
| Latest TCEC Superfinal | Won Season 29 vs Reckless, 59–41 (April 2026) | Lost Season 28 to Stockfish, 42.5–57.5 (2025); did not reach Season 29 final |
| TCEC Superfinal wins over the other | 19 TCEC titles in total as of Season 29 | Season 15 (2019), 53.5–46.5; Season 17 (2020), 52.5–47.5 |
| CCRL Blitz (Oct 17, 2026) | Stockfish 17.1, 1 CPU: 3771; Stockfish 19, 8 CPU: 3804 | Lc0 0.29.0, RTX 2080: 3696 |
Sources: design and history from Wikipedia: Stockfish and Wikipedia: Leela Chess Zero; TCEC results from Chessdom's reports on Season 28 and Season 29; ratings from the CCRL Blitz list.
Note what the TCEC Season 29 final tells you: the challenger was Reckless, another CPU alpha-beta engine, not Leela. At the very top, CPU engines currently hold the field.
CPU vs GPU engines: why both work
The two designs spend their time differently. Stockfish evaluates each position cheaply, so it can look at an enormous number of them and prune bad lines fast. Leela spends far more computation per position, because its network is big, and relies on that network's judgment to decide which few moves deserve a deeper look.
That is why Leela needs a GPU to be competitive and Stockfish does not. Both are now neural network engines. The difference is the size of the network and how much search sits on top of it. For how Stockfish's network works, see Stockfish NNUE.
Stockfish vs AlphaZero: the match that started it
Leela exists because of AlphaZero, DeepMind's self-taught engine.
- 2017 preprint (December 5, 2017): AlphaZero beat Stockfish 8 in 100 games, 28 wins, 72 draws, 0 losses, at a fixed one minute per move. Critics noted Stockfish had 64 threads but only a 1 GB hash, which Stockfish co-author Tord Romstad called far too small, and played without an opening book (Wikipedia: AlphaZero).
- 2018 Science paper (December 7, 2018): in a 1,000-game match at 3 hours + 15 seconds per move, AlphaZero scored 155 wins, 839 draws, 6 losses against Stockfish 8, with Stockfish on 44 CPU cores and a 32 GB hash. AlphaZero ran on four TPUs (Chess.com report, Science vol. 362, issue 6419, p. 1140).
AlphaZero was never released, so it never played a modern Stockfish. Leela rebuilt the approach in open source, and Stockfish answered by adding NNUE in 2020.
Our measurement: Stockfish in the browser
Chess Move Finder runs Stockfish 17.1 lite, single-threaded WebAssembly. To show what that means in practice, we ran the same build in Node on our server (Intel Xeon @ 2.10 GHz, 1 thread, 16 MB hash, fresh engine per run) with go movetime 1000 and go movetime 5000 on five positions, and recorded the final depth and nodes per second it reported.
| Position | Depth at 1s | Nodes/sec at 1s | Depth at 5s | Nodes/sec at 5s | Best move |
|---|---|---|---|---|---|
| Starting position | 23 | 795,192 | 29 | 709,106 | e4 |
| Opera Game after 10.Nxb5 | 22 | 750,262 | 27 | 764,230 | Qb4+ |
| Queen's Gambit Declined after 3...Nf6 | 23 | 736,312 | 27 | 636,470 | Bg5 |
| Rook ending, pawn on b7 | 23 | 937,178 | 26 | 1,044,232 | Rd1+ |
| King and pawn ending | 28 | 1,454,803 | 62 | 1,958,191 | Ke3 |
The test positions, so you can repeat it:
- Starting position:
rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq - 0 1 - Opera Game after 10.Nxb5:
rn2kb1r/p3qppp/2p2n2/1N2p1B1/2B1P3/1Q6/PPP2PPP/R3K2R b KQkq - 0 10 - Queen's Gambit Declined after 3...Nf6:
rnbqkb1r/ppp2ppp/4pn2/3p4/2PP4/2N5/PP2PPPP/R1BQKBNR w KQkq - 2 4 - Rook ending, pawn on b7:
1K1k4/1P6/8/8/8/8/r7/2R5 w - - 0 1 - King and pawn ending:
8/8/8/4k3/8/8/4PK2/8 w - - 0 1
Open the Opera Game position →
What it shows: one thread searches roughly 0.6–2 million positions per second, and reaches depth 22–23 in one second in normal middlegames. The near-empty king and pawn ending runs fastest and its depth balloons, because almost every line transposes. Your browser speed will differ by device; a desktop Stockfish on many cores is much faster. All five FENs were checked as legal with chess.js, and the Opera and QGD positions were generated from their move orders.
Which engine is best for analysis?
For everyday analysis, use Stockfish. It is the top-rated engine family on CCRL, runs on any CPU (including a phone browser), and gives fast, concrete answers about tactics, which is where most club games are decided. Leela is worth a second opinion in quiet strategic positions if you have a good GPU. In practice:
- Run the game through Game Review to find the moments that mattered.
- Open those moments in the analysis board and let it search for 5–10 seconds.
- Read the evaluation correctly with how to read chess engine evaluations.
- Get the plan in words from the coach, since neither engine explains itself.
Frequently asked questions
Is Leela stronger than Stockfish?
Not as of October 2026. Stockfish beat Leela 57.5–42.5 in the TCEC Season 28 Superfinal, and the best Lc0 on the CCRL Blitz list (3696) is below single-CPU Stockfish 17.1 (3771). Leela did win the TCEC Superfinals of Season 15 and Season 17.
Which engine is best for analysis?
Stockfish for most players: it is stronger on current lists, needs no GPU, and runs in the browser. Use Leela as a second opinion on strategic positions if you have the hardware.
Did AlphaZero beat Stockfish?
Yes. AlphaZero beat Stockfish 8 by 28 wins, 72 draws and no losses in the 2017 preprint, and by 155 wins, 839 draws and 6 losses in the 1,000-game match published in Science in December 2018. Both matches were against Stockfish 8, years before Stockfish added NNUE.
Is Stockfish a neural network engine?
Yes, since Stockfish 12 in 2020, and fully since July 2023, when the hand-crafted evaluation was removed. Its network is small and runs fast on CPUs, unlike Leela's.
For raw strength numbers against humans, see how strong is Stockfish. To test either style of thinking on your own positions, open the analysis board, and compare tools in chess AI tools.