About REMY
REMY is a chess engine written from scratch, one tested step at a time.
The name
REMY stands for Rapid Evaluation, Maximum Yield. It is also the engine's motto: REMY is tuned for blitz at 2 minutes plus 1 second per move, so every decision has to be made quickly, and the aim is to get the most out of each one. The name comes from the designer's daughter; more about the designer.
The goal
REMY's long-term aim is to beat Stockfish, the strongest chess engine in the world, at 2 minutes plus 1 second per move, on an ordinary desktop computer. On the CCRL Blitz list REMY's ratings use (computed October 17, 2026), the highest-rated Stockfish, Stockfish 19, is rated 3793 on one processor core and 3804 on eight.
That is a very long road. REMY's latest measured estimate is 2213 on the CCRL Blitz scale. Along the way there are milestones at 1500, 2000 and 2500: which have been passed, and what they mean.
REMY's goals
The main goal is to beat full-strength Stockfish: to score more than half the points against it in games of 2 minutes plus 1 second a move, on REMY's own Mac mini.
Two secondary goals
Lichess. A place in the top 10 of Lichess's bot leaderboard, in both bullet and blitz. On the way: a rating of 2,500 in bullet and in blitz, then 2,800, then the top 10. Other bots may run on much larger computers.
CCRL. A place in the top 50 of the CCRL Blitz list. On the way: a REMY release tested and listed by CCRL, on its Blitz and 40/15 lists; then 2,500 on CCRL Blitz; then 3,000; then 3,300; then the top 50. Only CCRL's own published figures count toward these. The estimates on this site follow REMY's progress between listings, but never count.
Why 2 minutes plus 1 second
Every rating here is on the CCRL Blitz scale, games of 2 minutes plus 1 second per move, because that is the time control the goal is defined at. Measuring every version on the same scale makes the climb comparable from one version to the next. How the clock is scaled to our test machine.
Ground rules
- REMY's training data comes only from its own games: it never trains on other engines' games, opening books or endgame databases.
- It contains no code from other chess engines or chess libraries: every rule, search and evaluation is written for REMY.
- It runs on an ordinary desktop computer: no graphics card and no cluster.
How REMY is built: one step at a time
REMY is made of separate parts. One knows the rules of chess, one remembers the game so far, one judges positions, one looks ahead, and one lets REMY talk to chess apps and other programs. Each new version makes one change, sometimes a small group of related ideas decided together, and nothing else.
A change is kept only if it passes two tests. First, it must work correctly: REMY's moves are checked against simpler, slower code that does the same job, and against published counts of every possible move sequence from standard test positions. Second, the new version must beat the version before it, in a match whose rules are fixed before it starts. Changes that fail are dropped, and every attempt, including failures, stays on record.
That is why REMY's history reads like a ladder: random moves, then counting material, then checking the reply, then skipping lines that cannot change its choice, then thinking deeper while time allows, then finishing exchanges before judging a position, then remembering positions it has already looked at, then learned knowledge of which squares suit each piece, then searching deeper by spending less time on moves that look weak, then the same search with its final learned square values, and now two more search ideas and a faster search that plays the same moves sooner.
The drawer: slide in an idea, let the games decide
REMY judges a position with a drawer of ideas. Each idea is one exact thing about a position that REMY can check, such as "a knight on this square" or "two bishops against a bishop and a knight".
The drawer is built so that trying a new idea is easy. Open the drawer, slide in an idea, and close it. The idea can come from a chess book (a principle from Nimzowitsch, say, spelled out exactly) or be a deliberately arbitrary hunch with no theory behind it at all. REMY then plays games against itself, and the idea's value is learned from REMY's own games: positive if it goes with winning, negative if it goes with losing, near zero if it makes no difference. Nobody sets the number by hand.
Then the idea has to earn its place. Three questions are kept apart: did the games give it a clear value; does that value still predict results in fresh games; and does REMY actually play better with it? Ideas also affect each other, so they are tried alone, together, and by taking each one out again. Sometimes the answer is "right track, wrong reason".
Every version of the drawer, including the ones that did not help, is kept in a permanent archive.
The first drawer, in REMY 7, holds one idea for every piece on every square, in the middlegame and in the endgame (programmers call these piece-square tables). Named ideas drawn from chess theory come in later stages.
See what REMY learned about each square.
Honesty rules
- Each version has to earn its place. A new version replaces the old one only after beating it in a match. Those results are shown as "self-play (against the previous version)".
- The ratings are REMY's own. REMY measures its rating by playing engines that already have CCRL Blitz ratings. CCRL has not tested REMY, so every rating is labelled "Estimated CCRL Blitz-scale rating (REMY measurement), not a CCRL rating".
- Every rating shows how sure it is. Each one comes with its likely range.
- The rules come before the games. How each match will be judged is written down before it starts, and failed or interrupted attempts stay on record.