Sites, apps and engines

Best Chess Engine: Strongest vs Most Human

Stockfish, Leela, Torch and Maia: which engine is strongest on the public rating lists, and which one plays like a person.

By the Chessage team · Published

Strongest: the rating lists

Stockfish 19 tops all three current lists, at 3650 on CCRL 40/15. The top of each list is crowded: engines within a few dozen points of each other trade places as new versions come out, so the number two spot moves around. Stockfish also won TCEC Season 29 (January to April 2026), beating Reckless in the finalsource.

FeatureCCRL 40/15CCRL BlitzCEGT best versions
Stockfish3650 (#1)source3804 (#1)source3650 (#1)source
Torch3640 (#7)source3778 (#3)source3620 (#6)source
Obsidian3636 (#11)source3774 (#4)source3598 (#21)source
Berserk3628 (#18)source3759 (#10)source3590 (#30)source
Dragon by Komodo3626 (#19–20)source3756 (#11–12)source3595 (#25)source
Leela Chess Zero3443 (#113–114)Lc0 0.28.0 w744204 64-bit 4CPU: CPU-only run of an old version: understates Leela on a GPUsource3696 (#45–46)Lc0 0.29.0 64-bit w753723 RTX2080: GPU run (RTX 2080); old versionsourceNot listed
Rating and rank on each list, best version of each engine. Lists differ in hardware and time control, so compare engines within a column, not across. Leela's entries are old versions; the CCRL 40/15 one ran without a GPU, which understates it. Between Stockfish and Torch on CCRL 40/15 sit Triumviratus, Gaia, Reckless, PlentyChess and pawnocchio.

Which ones you can use

  • Stockfish is open source (GPLv3)source and runs analysis on Lichess, Chess.com and many other sitessource. Version 19 came out on September 5, 2026source.
  • Leela Chess Zero is open source too and learned chess by playing itselfsource. It is at its best on a graphics card; the CPU-only build is much slowersource. It was Stockfish's TCEC runner-up five seasons in a rowsource.
  • Torch is Chess.com's engine. There is no public downloadsource; Chess.com members can pick it on the Analysis pagesource. Its developer said in 2024 that work on its strength had stoppedsource.

For checking your own games, any of them is far stronger than you need. What matters more is how the site explains the result. One concession: Chessage's review runs Stockfish 19's smaller "lite" network in your browser at depth 16, which is weaker than full Stockfish on a desktop. For deep opening or endgame work, use the full engine.

Most human: Maia

Strong engines make poor practice partners: they don't make the mistakes your opponents make. Maia, from the CSSLab research group, is trained to predict the move a person of a given rating would playsource. In the first Maia paper it matched the human move over 52% of the time at bestsource, while Stockfish matched 33–41%source. Maia-2 improved on that by almost two pointssource, and Maia-3 is built on a transformersource.

One catch: a Maia model plays stronger than the players it learned from, because it plays their most typical move every timesource. The Maia bots on Lichess show it:

FeatureTrained onRated on Lichess (rapid)
maia1Players around 11001494source
maia5Players around 15001603source
maia9Players around 19001710source
Lichess rapid ratings of the Maia bot accounts, read November 21, 2026.

Our test: do Maia bots blunder like people?

We use Maia-3 for the bots on Chessage's play page, so we measured it. On the same positions, Maia asked for a rating blundered about a quarter less often than real players of that rating, below 1800. At the top it stopped improving: players rated 2600+ blundered 3.4 times per 100 moves, Maia asked for 2600 blundered 5.4 times.

So we tuned each bot: it asks Maia for a lower rating, and from 1200 up a quick Stockfish check removes moves that players of that rating practically never make. After tuning, the bots blunder about as often as the people they imitate:

FeatureTuned botReal players
100022.622.6
120017.317.8
140014.915.1
160012.913.4
19008.810.1
21008.57.9
23004.95.1
25004.14.4
Blunders per 100 moves on the same positions (a blunder costs 20% or more of the winning chances), from 2026-10-17. Ratings are Lichess rapid. Players' numbers carry about ±1–2 of noise below 1800 and ±0.7–1.1 above.

The tuned bots also pick the exact move the real player chose 35% of the time at 1000, rising to about 57% at the top. Human-like doesn't mean predictable: people at the same rating disagree with each other too.

Which engine for which job

  • Checking your games: Stockfish, on any free review site. See free game review sites, tested.
  • Deep analysis or opening prep: full Stockfish on your own computer, or Leela if you have a good graphics card and want a second opinion.
  • Practice games: a human-like bot near your rating. Play a Maia bot on Chessage, or the Maia bots on Lichess.
  • Learning why a move is good: no engine explains itself. Use a review with a coach, then check the line with the engine.

Want to see the ratings across sites? Our Chess.com vs Lichess guide converts between them.

Frequently asked questions

What is the strongest chess engine?
Stockfish. Stockfish 19 is first on the CCRL 40/15, CCRL Blitz and CEGT rating lists as of November 2026, and it won TCEC Season 29 in April 2026. It is also free.
What engine does Chess.com use?
Chess.com's analysis page says it uses Stockfish, and Chess.com members can also pick Torch, Chess.com's own engine. We couldn't check which engine Game Review runs today without an account.
What is Maia chess?
Maia is a neural network from the CSSLab research group, trained to predict the move a human of a given rating would play, not the best move. That makes it a good sparring partner: it makes human mistakes.
Is Stockfish free?
Yes. Stockfish is open source under the GPLv3, and it is the engine behind analysis on Lichess, Chess.com and most chess apps.

Changelog

  • First published.