Leela Odds Leaderboard

Updated to make use of regular Contempt. These bots used nets meant for smaller handicaps, so they showed Leela as clearly losing in the initial position of those odds. Setting Contempt at levels that bring the evals to about zero gave gains in the twenty to thirty elo range, based on testing against simulated human bots. This should cut down the number of draws against human players who are not strong enough to score 50% at these handicaps. GM Deviatkin played many rook odds games at 5'3" both before and after, and the share of draws was much smaller after.

Menkib found and fixed the problem that was making the latest binary run super slow. The new version keeps all the improvements and actually runs slightly faster than the Nov 22 binary. It is now in all the bots. LeelaPieceOdds works with it too but still needs tuning.

The temperature parameter was reduced on all the bots, since it had been raised too much a couple of days earlier according to testing.

Updated to use the rook odds t3v3 net rather than a t1 net, and max nodes were reduced from 75k to 50k. The rook odds net tested clearly better here, because it thinks the opponent is a rook odds player, so it does not mind draws when down much more than a rook even with an anti draw setting.

Updated to use new binary by Borg that implements the idea of increasing max nodes when the position is no longer objectively hopeless. Some reductions were made to the basic max nodes numbers to compensate.

Updated to allow longer time controls, up to 45'+30" in general and up to 150'+60" for LKO and LRO. Also temperature settings were revised for all bots to reduce frequency of inferior lines while increasing maximum variety, should be a small elo gain.

Updated to use a new binary by @Menkib which implements Lucario6607's suggestion that Syzygy drawscores only be used when Leela's eval is worse than the drawscore. This in turn makes it safe enough to use Syzygy for draws even in 6 man (or 7 in the future) positions, as it will only seek them when it deems it more likely to lose than to win otherwise. Testing at Queen odds showed a gain of 27 elo over the previously best Syzygy implementation after about a thousand games for each version against three different human-like bots, on top of the 35 tested elo gain over no Syzygy. So the total tested gain for Syzygy is now 62 elo at queen odds.

Updated late Sept. 16 to use Syzygy for Leela only with a new binary by @Menkib. It considers only won positions for 6 man, both won and drawn ones for 3-4-5 man. I am re-testing vs. bots that need at least 10 nodes to be competitive, to avoid excessive repetition and stalemate draws. Early results confirm gains found in previous tests, tomorrow I'll report final results for queen odds. Gains are likely to be smaller for smaller odds, since the bot looks at more nodes and assumes better opponent play so is less likely to need TB correction, but there should still be a benefit.

Updated based on testing to set the value of TempVisitOffset to minus 6% of the node limit for that handicap.

New Lucario binary already running in LQO, together with small compensating reductions in the Search Contempt Limit. Testing showed 40 to 50 elo point gains against the human-based bots. Also a change to Policy temperature was reverted so that all use the default now.

Using new search-contempt binary by @amjsh, which boosts performance against humans by ~200 elo. This leaderboard shows all the games from this point on.

Updating the Bot to use specially trained queen-odds net (lqo_v2) by @Noah13.

Updated to make use of regular Contempt. These bots used nets meant for smaller handicaps, so they showed Leela as clearly losing in the initial position of those odds. Setting Contempt at levels that bring the evals to about zero gave gains in the twenty to thirty elo range, based on testing against simulated human bots. This should cut down the number of draws against human players who are not strong enough to score 50% at these handicaps. GM Deviatkin played many rook odds games at 5'3" both before and after, and the share of draws was much smaller after.

Menkib found and fixed the problem that was making the latest binary run super slow. The new version keeps all the improvements and actually runs slightly faster than the Nov 22 binary. It is now in all the bots. LeelaPieceOdds works with it too but still needs tuning.

Switched back to the Nov 22 Lc0 binary. The binaries of Feb 16 and 19 introduced new ideas and tested well on fixed nodes, but they were drastically slower than the Nov 23 2025 binary by Borg, by a factor of 6 with the LKO4 net. The switch solved the problems and both bots now run at great speed with tablebases. All the bots need retuning after this.

Updated with new parameter settings.

Updated with a parameter change, HybridMaxRatio raised from 0.8 to 0.9, which tested as a rather clear improvement.

The temperature parameter was reduced on all the bots, since it had been raised too much a couple of days earlier according to testing.

Upgraded based on bot testing. Max nodes were lowered, which showed an elo gain and also helps by reducing ponder time and the risk of overloading the computer, which can cause crashes and timeouts. The new binary that raises search contempt and max nodes once the position is no longer theoretically hopeless probably needs fewer nodes.

Updated again with changes to time management that speed up play at time controls where the node limits are not reached. It should help most in pure 1'0" bullet, where the bots sometimes forfeited on time or dropped below the overhead buffer and ended up playing pure policy moves. Total tested gain at 1'0" for all the changes of the past week is 90 elo at rook odds.

The single piece bots were upgraded over the past couple of days to use a new binary that combines changes by Menkib and Lucario. The Menkib changes reduce the VRAM requirement, which allowed for higher max Concurrency and for putting knight odds back on the 5090. Menkib also changed how temperature works so that Smart Pruning can be used again without causing missed mates, which should speed up play. The Lucario change applies the increase in max nodes when the score suggests Leela is no longer clearly losing to the Search Contempt and the related Scale Factor, so that these values stay the same percentage of max nodes. All of this still needs tuning, which is ongoing.

Upgraded to a higher node count, maybe a 15 elo gain, and the opening book was updated as well.

Based on new testing the node goal was raised from 50k to 70k.

Updated to use new binary by Borg that implements the idea of increasing max nodes when the position is no longer objectively hopeless. Some reductions were made to the basic max nodes numbers to compensate.

Updated to allow longer time controls, up to 45'+30" in general and up to 150'+60" for LKO and LRO. Also temperature settings were revised for all bots to reduce frequency of inferior lines while increasing maximum variety, should be a small elo gain.

Updated to use a new binary by @Menkib which implements Lucario6607's suggestion that Syzygy drawscores only be used when Leela's eval is worse than the drawscore. This in turn makes it safe enough to use Syzygy for draws even in 6 man (or 7 in the future) positions, as it will only seek them when it deems it more likely to lose than to win otherwise. Testing at Queen odds showed a gain of 27 elo over the previously best Syzygy implementation after about a thousand games for each version against three different human-like bots, on top of the 35 tested elo gain over no Syzygy. So the total tested gain for Syzygy is now 62 elo at queen odds.

Updated late Sept. 16 to use Syzygy for Leela only with a new binary by @Menkib. It considers only won positions for 6 man, both won and drawn ones for 3-4-5 man. I am re-testing vs. bots that need at least 10 nodes to be competitive, to avoid excessive repetition and stalemate draws. Early results confirm gains found in previous tests, tomorrow I'll report final results for queen odds. Gains are likely to be smaller for smaller odds, since the bot looks at more nodes and assumes better opponent play so is less likely to need TB correction, but there should still be a benefit.

Updated with new search parameter values which showed a ten elo gain in testing.

Updated based on testing to set the value of TempVisitOffset to minus 6% of the node limit for that handicap.

New binary by Lucario6607 as is already in LeelaQueenOdds, with a gradual transition between the two search methods using a logarithmic formula. The parameter values chosen for LRO yielded a similar 20+ elo gain in testing to that observed with LQO.

New Lucario binary already running in LQO, together with small compensating reductions in the Search Contempt Limit. Testing showed 40 to 50 elo point gains against the human-based bots. Also a change to Policy temperature was reverted so that all use the default now.

Using new search-contempt binary by @amjsh, which boosts performance against humans by ~200 elo. This leaderboard shows all the games from this point on.

Update with new special rook odds net, node limit of 8000.

Menkib found and fixed the problem that was making the latest binary run super slow. The new version keeps all the improvements and actually runs slightly faster than the Nov 22 binary. It is now in all the bots. LeelaPieceOdds works with it too but still needs tuning.

Switched back to the Nov 22 Lc0 binary. The binaries of Feb 16 and 19 introduced new ideas and tested well on fixed nodes, but they were drastically slower than the Nov 23 2025 binary by Borg, by a factor of 6 with the LKO4 net. The switch solved the problems and both bots now run at great speed with tablebases. All the bots need retuning after this.

Updated with new parameters. The sc limit and a related parameter were set too high.

The temperature parameter was reduced on all the bots, since it had been raised too much a couple of days earlier according to testing.

Upgraded based on bot testing. Max nodes were lowered, which showed an elo gain and also helps by reducing ponder time and the risk of overloading the computer, which can cause crashes and timeouts. The new binary that raises search contempt and max nodes once the position is no longer theoretically hopeless probably needs fewer nodes.

Updated again with changes to time management that speed up play at time controls where the node limits are not reached. It should help most in pure 1'0" bullet, where the bots sometimes forfeited on time or dropped below the overhead buffer and ended up playing pure policy moves. Total tested gain at 1'0" for all the changes of the past week is 90 elo at rook odds.

The single piece bots were upgraded over the past couple of days to use a new binary that combines changes by Menkib and Lucario. The Menkib changes reduce the VRAM requirement, which allowed for higher max Concurrency and for putting knight odds back on the 5090. Menkib also changed how temperature works so that Smart Pruning can be used again without causing missed mates, which should speed up play. The Lucario change applies the increase in max nodes when the score suggests Leela is no longer clearly losing to the Search Contempt and the related Scale Factor, so that these values stay the same percentage of max nodes. All of this still needs tuning, which is ongoing.

Updated with a 0.1 adjustment to DrawScore in the direction of avoiding draws.

No upgrade to the engine settings yet, but the opening book was updated.

Updated to use new binary by Borg that implements the idea of increasing max nodes when the position is no longer objectively hopeless. Some reductions were made to the basic max nodes numbers to compensate.

Updated to allow longer time controls, up to 45'+30" in general and up to 150'+60" for LKO and LRO. Also temperature settings were revised for all bots to reduce frequency of inferior lines while increasing maximum variety, should be a small elo gain.

Updated to use a new binary by @Menkib which implements Lucario6607's suggestion that Syzygy drawscores only be used when Leela's eval is worse than the drawscore. This in turn makes it safe enough to use Syzygy for draws even in 6 man (or 7 in the future) positions, as it will only seek them when it deems it more likely to lose than to win otherwise. Testing at Queen odds showed a gain of 27 elo over the previously best Syzygy implementation after about a thousand games for each version against three different human-like bots, on top of the 35 tested elo gain over no Syzygy. So the total tested gain for Syzygy is now 62 elo at queen odds.

Updated late Sept. 16 to use Syzygy for Leela only with a new binary by @Menkib. It considers only won positions for 6 man, both won and drawn ones for 3-4-5 man. I am re-testing vs. bots that need at least 10 nodes to be competitive, to avoid excessive repetition and stalemate draws. Early results confirm gains found in previous tests, tomorrow I'll report final results for queen odds. Gains are likely to be smaller for smaller odds, since the bot looks at more nodes and assumes better opponent play so is less likely to need TB correction, but there should still be a benefit.

Updated to use the same new binary by Lucario6607 as is already in LeelaQueenOdds and LeelaRookOdds, with a gradual transition between the two search methods using a logarithmic formula. The parameter values chosen for LKO yielded a 20 elo gain in testing, similar to that observed with LQO.

Updated based on testing to set the value of TempVisitOffset to minus 6% of the node limit for that handicap.

Reduction of the Search Contempt Limit from 520 to 500 based on bot testing.

New Lucario binary already running in LQO, together with small compensating reductions in the Search Contempt Limit. Testing showed 40 to 50 elo point gains against the human-based bots. Also a change to Policy temperature was reverted so that all use the default now.

Using new search-contempt binary by @amjsh, which boosts performance against humans by ~200 elo. This leaderboard shows all the games from this point on.

Update the bot to the latest knight-odds net lko_v3a.