I toast to LeelaqueenOdds and its creators. I raise my wine glass high and shout: Long live the queen. We have a queen!!

[Father]

The only way I can think of for now for me to aspire to occupy top spots on the LeelaQueenOdds robot leaderboard would be for the robot to acquire new strengths that make it much more lethal. I'd like to ask something: Can LeelaQueenOdds substantially increase its playing strength against the best in the bullet leaderboard to the point of hitting them really hard and strong? If so, I believe it would benefit me greatly. Thanks again.]

LeelaQueenOdds was upgraded late night Aug. 18 to a new neural net, trained by @Ira_Jed, based on 350,000 queen odds games against human-simulating bots, much more than ever before used. My testing showed it to perform 73 elo better giving queen odds than the net previously used. Results vs humans do already seem better. We'll have to see if the improvement is enough to make a difference against the superstar players who top the Leaderboard. LeelaPieceOdds was upgraded this morning to use this same net for all the odds of queen plus anything, as testing showed that it improved over the previously used net by a substantial margin in most cases. It was also upgraded to use the same new search (by Lucario6607) as the other bots are now using, which boosted them all by forty to fifty elo. I was hoping that the new net would also be better at queen for knight odds, but testing did not prove that to be so, so no change to that bot.

Thank you very much Mr. Larry Kaufman for your information. I would like to ask you, given the current strength of the four robots in FIDE ratings, what would you believe the tentative FIDE rating of the four robots could be in time controls of: 3 plus 2; 3-0; 1 plus 1; 1-0; 5-0; 10-0; 20-0, and in the regulatory FIDE time used in official tournaments for the World Human Championship. Thank you Mr. Larry Kaufman. Note: Personally, I have noticed, since yesterday and today, that the robots Knight and Queen are stronger. I don't know if it's just my imagination, but that's my perception. Taking a simple draw against LeelaKnight is costing me a lot of difficulty, even though I vary my positions.

Well, as I said LeelaQueenOdds did receive a major upgrade late on Aug 18, so it's not surprising you see a difference. Just a few hours ago the LeelaPieceOdds bot was upgraded across the board; the new search was added, and all the multi-piece handicaps were given new nets and settings, in some cases with a massive gain in strength (mostly at the huge handicaps). Knight odds and Rook odds still use the same net as they did at the end of 2024, but they do use the new search which might be a 50 elo boost, so that is probably what you are observing at knight odds. The switch to a 5090 and to using onnx-trt back in June also boosted all the bots, settings have been refined, and the opening books are updated often. These things add up, and results just keep getting better. It's too soon to know the full effect of these recent changes. Probably LeelaQueenOdds is around FIDE 2400 now at 3'2", meaning that is the rating needed to break even. But note that doesn't mean it could get a 2400 rating against much lower rated players if they play for a draw (as you do), because Leela isn't programmed to avoid draws unless you have made a significant error, she assumes you are strong enough to score 50% at the odds. Not enough data for LeelaQueenforKnight to be very precise, probably adding 300 to the queen odds rating is about right. Leela rook odds is clearly above human level at 3'2", perhaps about 2950 FIDE now. Leela Knight odds is perhaps 3100 FIDE at 3'2" blitz if playing against 2800 rated players. But if it had to play FIDE 2000s every game, and they all played for draws, it might only get 2400 or so. Perhaps the best way to think of these estimated ratings is to imagine that all games were played with a rule that in the event of a draw, the game would be replayed until someone won, with the draws not counting (19th century rules!). As for other time limits, the adjustments I gave a few months ago should be the same. At classical time limits, we only know that we performed about 2425 FIDE at knight odds vs Joel Benjamin in January, but improvements are enough that I believe we would do better than that today even at rook odds, much better at knight, so maybe 2450 FIDE at rook, 2600 at knight. Perhaps 1900 at queen?

[pgn][Event "casual bullet game"]
[Site "https://lichess.org/CVB8Vayl"]
[Date "2025.08.22"]
[White "Catecan"]
[Black "LeelaKnightOdds"]
[Result "1/2-1/2"]
[GameId "CVB8Vayl"]
[UTCDate "2025.08.22"]
[UTCTime "14:50:44"]
[WhiteElo "2022"]
[BlackElo "2000"]
[BlackTitle "BOT"]
[Variant "From Position"]
[TimeControl "60+0"]
[ECO "?"]
[Opening "?"]
[Termination "Normal"]
[FEN "r1bqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq - 0 1"]
[SetUp "1"]
[Annotator "lichess.org"]

1. d4 e6 2. f4 Nf6 3. e3 c5 4. Nf3 a6 5. c3 Qc7 6. Bd3 b5 7. Qe2 Bb7 8. O-O Be7 9. a3 c4 10. Bc2 O-O 11. Nbd2 d5 12. g3 a5 13. h4 Ne8 14. Rf2 b4 15. Rg2 Nd6 16. g4 f6 17. g5 f5 18. h5 b3 19. Bd1 Rfb8 20. h6 g6 21. Nb1 Bc6 22. Kh1 a4 23. Kg1 Ne4 24. Kh1 Bd6 25. Kg1 Be7 26. Kh1 Nd6 27. Kg1 Nb5 28. Kh1 Bd6 29. Kg1 Qe7 30. Kh1 Rb7 31. Kg1 Be8 32. Kh1 Rab8 33. Kg1 Bc7 34. Kh1 Ba5 35. Kg1 Nd6 36. Qe1 Nb5 37. Kh1 Nd6 38. Kg1 Rc8 39. Kh1 Rcb8 40. Kg1 { The game is a draw. } 1/2-1/2[/pgn]

[Father]

The only way I can think of for now for me to aspire to occupy top spots on the LeelaQueenOdds robot leaderboard would be for the robot to acquire new strengths that make it much more lethal. I'd like to ask something: Can LeelaQueenOdds substantially increase its playing strength against the best in the bullet leaderboard to the point of hitting them really hard and strong? If so, I believe it would benefit me greatly. Thanks again.]

LeelaQueenOdds was upgraded late night Aug. 18 to a new neural net, trained by @Ira_Jed, based on 350,000 queen odds games against human-simulating bots, much more than ever before used. My testing showed it to perform 73 elo better giving queen odds than the net previously used. Results vs humans do already seem better. We'll have to see if the improvement is enough to make a difference against the superstar players who top the Leaderboard. LeelaPieceOdds was upgraded this morning to use this same net for all the odds of queen plus anything, as testing showed that it improved over the previously used net by a substantial margin in most cases. It was also upgraded to use the same new search (by Lucario6607) as the other bots are now using, which boosted them all by forty to fifty elo. I was hoping that the new net would also be better at queen for knight odds, but testing did not prove that to be so, so no change to that bot.

Thank you very much Mr. Larry Kaufman for your information. I would like to ask you, given the current strength of the four robots in FIDE ratings, what would you believe the tentative FIDE rating of the four robots could be in time controls of: 3 plus 2; 3-0; 1 plus 1; 1-0; 5-0; 10-0; 20-0, and in the regulatory FIDE time used in official tournaments for the World Human Championship. Thank you Mr. Larry Kaufman. Note: Personally, I have noticed, since yesterday and today, that the robots Knight and Queen are stronger. I don't know if it's just my imagination, but that's my perception. Taking a simple draw against LeelaKnight is costing me a lot of difficulty, even though I vary my positions.

Well, as I said LeelaQueenOdds did receive a major upgrade late on Aug 18, so it's not surprising you see a difference. Just a few hours ago the LeelaPieceOdds bot was upgraded across the board; the new search was added, and all the multi-piece handicaps were given new nets and settings, in some cases with a massive gain in strength (mostly at the huge handicaps). Knight odds and Rook odds still use the same net as they did at the end of 2024, but they do use the new search which might be a 50 elo boost, so that is probably what you are observing at knight odds. The switch to a 5090 and to using onnx-trt back in June also boosted all the bots, settings have been refined, and the opening books are updated often. These things add up, and results just keep getting better. It's too soon to know the full effect of these recent changes. Probably LeelaQueenOdds is around FIDE 2400 now at 3'2", meaning that is the rating needed to break even. But note that doesn't mean it could get a 2400 rating against much lower rated players if they play for a draw (as you do), because Leela isn't programmed to avoid draws unless you have made a significant error, she assumes you are strong enough to score 50% at the odds. Not enough data for LeelaQueenforKnight to be very precise, probably adding 300 to the queen odds rating is about right. Leela rook odds is clearly above human level at 3'2", perhaps about 2950 FIDE now. Leela Knight odds is perhaps 3100 FIDE at 3'2" blitz if playing against 2800 rated players. But if it had to play FIDE 2000s every game, and they all played for draws, it might only get 2400 or so. Perhaps the best way to think of these estimated ratings is to imagine that all games were played with a rule that in the event of a draw, the game would be replayed until someone won, with the draws not counting (19th century rules!). As for other time limits, the adjustments I gave a few months ago should be the same. At classical time limits, we only know that we performed about 2425 FIDE at knight odds vs Joel Benjamin in January, but improvements are enough that I believe we would do better than that today even at rook odds, much better at knight, so maybe 2450 FIDE at rook, 2600 at knight. Perhaps 1900 at queen?

Good evening, Mr. Larry Kaufman. In my own experience, I've observed, and above all, "perceived," a marvelous development in robots. I've learned to stay calm and maintain my composure during one-minute games against robots, allowing me to make quick, in-depth strategic analyses. What I don't understand is why it's not possible for me to achieve a 100% draw against LeelaQueenOdds. Not only is that not the case, but the computer often traps me. LeelaQueeOddds is truly very strong, and the difficulty of achieving a draw requires me to use all my strength. All these robots are marvels of the human soul. Congratulations.

[Father]

The only way I can think of for now for me to aspire to occupy top spots on the LeelaQueenOdds robot leaderboard would be for the robot to acquire new strengths that make it much more lethal. I'd like to ask something: Can LeelaQueenOdds substantially increase its playing strength against the best in the bullet leaderboard to the point of hitting them really hard and strong? If so, I believe it would benefit me greatly. Thanks again.]

LeelaQueenOdds was upgraded late night Aug. 18 to a new neural net, trained by @Ira_Jed, based on 350,000 queen odds games against human-simulating bots, much more than ever before used. My testing showed it to perform 73 elo better giving queen odds than the net previously used. Results vs humans do already seem better. We'll have to see if the improvement is enough to make a difference against the superstar players who top the Leaderboard. LeelaPieceOdds was upgraded this morning to use this same net for all the odds of queen plus anything, as testing showed that it improved over the previously used net by a substantial margin in most cases. It was also upgraded to use the same new search (by Lucario6607) as the other bots are now using, which boosted them all by forty to fifty elo. I was hoping that the new net would also be better at queen for knight odds, but testing did not prove that to be so, so no change to that bot.

Thank you very much Mr. Larry Kaufman for your information. I would like to ask you, given the current strength of the four robots in FIDE ratings, what would you believe the tentative FIDE rating of the four robots could be in time controls of: 3 plus 2; 3-0; 1 plus 1; 1-0; 5-0; 10-0; 20-0, and in the regulatory FIDE time used in official tournaments for the World Human Championship. Thank you Mr. Larry Kaufman. Note: Personally, I have noticed, since yesterday and today, that the robots Knight and Queen are stronger. I don't know if it's just my imagination, but that's my perception. Taking a simple draw against LeelaKnight is costing me a lot of difficulty, even though I vary my positions.

Well, as I said LeelaQueenOdds did receive a major upgrade late on Aug 18, so it's not surprising you see a difference. Just a few hours ago the LeelaPieceOdds bot was upgraded across the board; the new search was added, and all the multi-piece handicaps were given new nets and settings, in some cases with a massive gain in strength (mostly at the huge handicaps). Knight odds and Rook odds still use the same net as they did at the end of 2024, but they do use the new search which might be a 50 elo boost, so that is probably what you are observing at knight odds. The switch to a 5090 and to using onnx-trt back in June also boosted all the bots, settings have been refined, and the opening books are updated often. These things add up, and results just keep getting better. It's too soon to know the full effect of these recent changes. Probably LeelaQueenOdds is around FIDE 2400 now at 3'2", meaning that is the rating needed to break even. But note that doesn't mean it could get a 2400 rating against much lower rated players if they play for a draw (as you do), because Leela isn't programmed to avoid draws unless you have made a significant error, she assumes you are strong enough to score 50% at the odds. Not enough data for LeelaQueenforKnight to be very precise, probably adding 300 to the queen odds rating is about right. Leela rook odds is clearly above human level at 3'2", perhaps about 2950 FIDE now. Leela Knight odds is perhaps 3100 FIDE at 3'2" blitz if playing against 2800 rated players. But if it had to play FIDE 2000s every game, and they all played for draws, it might only get 2400 or so. Perhaps the best way to think of these estimated ratings is to imagine that all games were played with a rule that in the event of a draw, the game would be replayed until someone won, with the draws not counting (19th century rules!). As for other time limits, the adjustments I gave a few months ago should be the same. At classical time limits, we only know that we performed about 2425 FIDE at knight odds vs Joel Benjamin in January, but improvements are enough that I believe we would do better than that today even at rook odds, much better at knight, so maybe 2450 FIDE at rook, 2600 at knight. Perhaps 1900 at queen?

** LeelaKnight is truly very strong, and the difficulty of achieving a draw requires me to use all my strength. All these robots are marvels of the human soul. Congratulations.

[M ANSARI]

I see that I did not find instructions how to download leelaqueen odds in discord.

searching in google lead me to the following link and I am not sure if it is the LeelaQueenOdds or some old version

https://github.com/notune/LeelaQueenOdds

I see release date is 13.11.2024

Yes, that is the net used by LeelaQueenOdds from last November up until one day ago. We upgraded (?) to a new net also trained for queen odds, by "@Ira_Jed and @Lucario6607. It's not yet clear that it is stronger, but I think on balance the evidence suggests that it is a bit stronger against both human-like bots and against actual humans. Right after the switch, it played four blitz (3'0") games against GM Jaan Ehlvest, who was World number 5 about 35 years ago and is still, at age 63, rated well over the GM threshold at 2532 FIDE. Leela won three and lost one. No games decided by time, but Ehlvest did move a bit faster than necessary for a 3' game, more as if they were 2' games. He said that he thinks this was the main reason for his loss, but even so winning a blitz match at queen odds from a former top-five player indicates that the new net is not bad! By a curious coincidence, Ehlvest is the strongest player I ever defeated in a tournament game, around 1996 I believe, in the World Open.

I am actually playing a lot of games against the QueensOdds bot and find it quite challenging even at 5min +3s. I find it really annoying how it can avoid exchanges and still somehow manage to keep pieces in positionally viable! I do feel that it plays incredibly inconsistent in strength and that would be the main thing I find that should be fixed. I have a feeling that sometimes it is playing with hardware shared with many players. I have recently upgraded my computer and it has the Ultra i9 285 processor and an RTX 5080 card. Is there a way to create my own bot to play with at home. I know that the bot uses a 5090 card but I think that maybe a non shared 5080 would play stronger maybe? I decided not to get the 5090 as it has a huge thermal and voltage footprint and would need active cooling which could end up being noisy.

[lkaufman]

I see that I did not find instructions how to download leelaqueen odds in discord.

searching in google lead me to the following link and I am not sure if it is the LeelaQueenOdds or some old version

https://github.com/notune/LeelaQueenOdds

I see release date is 13.11.2024

Yes, that is the net used by LeelaQueenOdds from last November up until one day ago. We upgraded (?) to a new net also trained for queen odds, by "@Ira_Jed and @Lucario6607. It's not yet clear that it is stronger, but I think on balance the evidence suggests that it is a bit stronger against both human-like bots and against actual humans. Right after the switch, it played four blitz (3'0") games against GM Jaan Ehlvest, who was World number 5 about 35 years ago and is still, at age 63, rated well over the GM threshold at 2532 FIDE. Leela won three and lost one. No games decided by time, but Ehlvest did move a bit faster than necessary for a 3' game, more as if they were 2' games. He said that he thinks this was the main reason for his loss, but even so winning a blitz match at queen odds from a former top-five player indicates that the new net is not bad! By a curious coincidence, Ehlvest is the strongest player I ever defeated in a tournament game, around 1996 I believe, in the World Open.

I am actually playing a lot of games against the QueensOdds bot and find it quite challenging even at 5min +3s. I find it really annoying how it can avoid exchanges and still somehow manage to keep pieces in positionally viable! I do feel that it plays incredibly inconsistent in strength and that would be the main thing I find that should be fixed. I have a feeling that sometimes it is playing with hardware shared with many players. I have recently upgraded my computer and it has the Ultra i9 285 processor and an RTX 5080 card. Is there a way to create my own bot to play with at home. I know that the bot uses a 5090 card but I think that maybe a non shared 5080 would play stronger maybe? I decided not to get the 5090 as it has a huge thermal and voltage footprint and would need active cooling which could end up being noisy.

The new net in the bot is not currently publicly available, but the one that was used until about a month ago is available, and it wasn't too much weaker. So you could make your own bot. The queen odds bot is on a 4090, the other four bots are on the 5090. The sharing among all the LQO players should not affect the moves played at 5'3", only the time it takes to make the move, because LQO currently has a 20k node limit which should be easily reached at that time control even with 12 players playing at once. Going beyond 20k nodes is not helpful with the current bot, perhaps that may be different soon. So for now playing on your own hardware wouldn't do much other than speed up the bot, which reduces your own ponder time. As for inconsistency, it would be nice to fix this, but it may simply be due to the bot having to "guess" what the human is capable of seeing, which is not easy to do! But I think some things we are working on will at least help somewhat.