Leela Odds Challenges on Lichess: Queen for knight & double knight odds

It has been almost exactly 2 years since the Leela odds match with GM David Navara, and a lot has happened since. To showcase this progress, selected odds setups have been chosen for public blitz challenges: Queen for Knight (already finished) and Double Knight Odds (starting now).

Besides the four featured Leela odds bots on LiChess, there is a fifth one, LeelaQueenForKnight, which gives odds of a queen in return for the opponent’s queenside knight. It has received little attention until now, but I decided to remedy this by offering a modest $300 prize for the first titled identified player with an established Lichess blitz rating over 2500 to make a plus score in 30 consecutive games against it within a 16 hour window, Leela playing White as is standard for piece odds play. The prize would not be awarded to anyone suspected of using engine help, but I’m happy to report that there were no participants suspected of this. The bots run on my home computers, with queen odds, rook odds, and knight odds on a 5090 and LeelaPieceOdds plus LeelaQueenForKnight running on a 4090 with up to eight total games at once on the 4090 (22 on the 5090).

Results of the Queen for Knight blitz challenge

Time limit for the blitz challenge was initially set at 3'0", with 2’+1" or 1’+2" also allowed. Leela generally averaged between 1 and 2 seconds per move. Note that unlike the single piece odds (and two knights), the LeelaQueenForKnight bot doesn’t have its own net, it used a rook odds net by “Naca Fox”. Quite a few IMs and a GM played it the first day, but the results against these strong players were so lopsided (over 86% after 300 games) that I raised the allowed max time limit to 3’+1" blitz, which almost everyone switched to after that.

Eventually more than 400 games were played by IMs and GMs (plus NM “Chess-Network” whose online ratings are all at IM level) at the longer 3’+1" time control. It still took several days until someone finally won the prize, despite several GMs giving it a try without success, namely Hovhannes Gabuzyan, Max Warmerdam, and Jorden van Foreest (3 games only) at the faster time control, and Gabuzyan, Keith Arkell, and Jeffery Xiong at the 3’+1" time control. None of these GMs nor the nine IMs or “Chess-Network” (or any FM) even came close to an even score at 3’+1", except for the ultimate winner, Super-GM Jeffery Xiong (over 2700 FIDE blitz, over 2850 Lichess blitz, and over 3200(!) chess.com blitz rating), with a total score of 20 to 14, despite a losing record on the first day he tried, and as a runner-up IM Felix Xie with a 25-32 score at 3’+1", way better than anyone else except Xiong. IMs Vincent Rothuis and Kacper Drozdowski get honorable mentions for scoring around or above 30% at 3’+1", while no one else even came close to that!

The performance ratings for the bot (at 3’+1" only) were about: FIDE classical 2614, FIDE blitz 2582, Lichess Blitz 2860, and Chess.com Blitz 3036. But even players rated above these numbers failed to make an even score, except for winner Xiong. The performance at the faster time controls was over a hundred Elo higher in each category, although it should be noted that most of the slower games were played after gaining experience at the faster time control. The winner commented that prior experience playing many 1’+0" bullet games with the full queen odds bots was quite helpful for this challenge. Another observation from participants was that it was fairly easy to make draws since Leela would generally allow a repetition if still far behind, but that wouldn’t generally help them to win the prize.

Insights from the Queen for Knight challenge

As for the games themselves, of course there was huge variety, both in openings and in the nature of the games. In general Leela would often aim for a closed position where it was hard to see how to win without giving back material despite the huge queen for knight gap, and often players would fail to win endgames if they were “only” up by a piece or so! It generally took more than 70 moves on average to defeat Leela, so it is quite understandable that having an increment of even one second helped a great deal. I think Leela showed just how many obstacles can be thrown in the way if you play well enough and don’t lose hope despite starting with a resignable position. But time does play a major role; I’m sure Leela’s results would decay noticeably with each added minute of base time or second of increment. Since most blitz games with the odds bots are played at 3'2", with many at 5'3", I think a fair overall Lichess blitz rating for this bot would be about 2800. Strong GMs like Xiong can beat it in blitz, but not “par” GMs.

There are too many people who have contributed to this odds project to name them all, but thanks to all for your contributions. My own role, aside from providing the hardware, is to update the opening book, to select which nets for each bot to use, and to tune parameters. The actual programming and net training is a combined effort from several different people, independent of the regular Leela project though starting from standard Leela engine and nets.

Coming up next: Double Knight odds blitz challenge

For the next challenge, the odds will switch to two (white) knights against the LeelaPieceOdds bot. It should be a slightly larger handicap than queen for knight, but because it has its own net, actual results so far against humans have been almost as impressive as in the Q for N challenge, although 2900+ (Lichess blitz) IM Kacper Drozdowski did beat it convincingly a few days ago in test games at 3’+1", which he was not able to achieve at queen for knight odds. Therefore we will solely use the 3’+0" blitz time control as we did at first with queen for knight, and only increase it to 3’+1" after several strong top blitz players like Drozdowski and Xiong fail to claim the prize at 3’+0".

Note that unlike in the first challenge, 2’+1" will not be treated as an equivalent replacement for 3’+0". Other than that, the format will be mostly the same: $300 prize, same player restrictions (titles below FM or Lichess blitz rating below 2500 may be allowed on request for streamers or players with high chess.com ratings by contacting Hissha via Lichess or in the Lc0 Discord), but some changes to the winning conditions. This time the prize will go to the first player to make an overall plus score (once the challenge is live) which includes at least ten wins (again assuming no evidence of cheating). Games won due to engine disconnection/crash/engine stalling out don’t count.

While theoretically one could win the prize in ten games, I don’t expect any human to do that, and it likely will again take a “super” GM (one who has exceeded 2700 FIDE) to win the prize, or at least someone at that level in blitz (2900 Lichess blitz), but I could be wrong. In general I think games at two knights odds tend to be more exciting than at queen for knight, as the bot is more likely to take risks trying to win when it has a queen.

Summary of the challenge conditions:

• LeelaPieceOdds plays white with both of its knights removed lc0.org/NN
• time control is 3’+0"
• challenge is open to all titled players FM and above (GM, IM, FM) with a 2500+ Lichess blitz rating (50+ games) and their identity revealed, with possible extensions after personal request (see above)
• the first participant who makes an overall plus score which includes at least ten wins (starting now, i.e. 2026-03-21 00:00 UTC) will win the challenge and the $300 prize