It could have been an 11-way tie for first, had Halogen and Leela split the points. However, Leela managed to outplay Halogen in a QGD Slav and jumped ahead of competition.

The standings bring exciting matchups in the next round , where we will see Leela Chess Zero – Stockfish and Ceres – RofChade

Follow TCEC live: Official website / TCEC Chess TV / Lichess / Chessdom live More: Interview with Gabriele Lombardo from Obsidian

Standings

1. Leela Chess Zero 4,5/6

2-10. Ethereal, Stockfish, Ceres, Ginkgo, rofChade, Stormphrax, Viridithas, Devre, and Stoofvlees 4,0/6

11-17. Halogen, Velvet, Berserk, Minic, Obsidian, Caissa and Igel 3,5/6

Total 44 engines, full standings here

See Obsidian in action now live at https://tcec-chess.com/

Obsidian is currently one of the top engines in the world. How did you decide to enter computer chess programming?

My interest in chess engines began about 2 years ago. I began playing chess with a friend OTB, and soon I found out about chess.com and began playing there. In chess.com the game review feature sparked my interest in chess engines because I was wondering, “How does this guy know what moves are good and bad?”. So I found out about chess engines and I worked for some months on a kind of “toy” engine. Then, I decided to rewrite everything from scratch in a more serious manner and that was what is now Obsidian.

The other day in chat someone mentioned that you are very young, actually the youngest of all developers in TCEC

That’s right. Every engine developer I have ever talked to is older than me. I am 16 years old and will turn 17 in March 2025.

The new season of Top Chess Engine Championship TCEC has started with Swiss 7, the strongest chess event regarding playing strength. What are your expectations?

Obsidian most recent versions defeated Komodo Dragon 3.3 (the latest version) on SP-CC and Ipmanchess (two leaderboards). I hope Obsidian can repeat this performance in the Top Chess Engine Championship. The number of games played on TCEC is limited, so bad luck could happen. Hopefully not!

What do you consider the biggest strength of Obsidian? And the biggest weakness?

The biggest strength of Obsidian is the evaluation. It is not a random Leela data NNUE – my NNUE is, by a big margin, stronger than the NNUE of any other Leela data engine, outside of Stockfish of course. Excluding the datagen process (which is missing), there is a lot of effort into training. It was speculated about 2 months ago by a Torch developer that Obsidian’s evaluation might be almost as strong as that of Torch.

The biggest weakness is probably the search algorithm, being the engine very fast in terms of nps.

What are the original inventions in Obsidian?

I don’t know if calling them “inventions” is appropriate. There are several original implementation details that gained Elo in my engine in comparison to the way I’ve seen other engines do things. They are very technical and specific things I won’t go over.

Regarding speed, Obsidian until recently had 2 completely original techniques about efficient updates of the neural network. Recently I had to remove them due to their high RAM consumption – but they were significant speedups.

Lastly, I don’t do datagen, but in the training process itself I use various techniques that result in a (far) stronger evaluation than that of engines with similar training data.

What are the things that you are currently researching?

In this period I am trying to introduce 3 hidden layers (instead of 1) in the neural network evaluation of Obsidian. Every attempt until now has been a failure. I can successfully measure an increased strength of evaluation, but the speed loss is exaggerate.

Does Obsidian have plans to pull away from using Leela data for NNUE towards a more original approach?

Generating data by myself would be impossible for now, due to my very restricted amount of resources/hardware. I don’t think I could find someone else who would dedicate to generate data for Obsidian either.

Make a prediction for top 5 in the current TCEC Swiss 7

My prediction is: Stockfish, LCZero, Ceres, Berserk, Obsidian

Do you follow human chess games? Are you going to follow the Chess Olympiad in Budapest?

I watch human games sometimes. I mostly follow Hikaru, and last week I’ve been watching the Speed Chess Championship. I didn’t know about the chess Olympiad in Budapest until I read your message – I will probably watch some games.

TCEC Swiss 7 will be the first event of the season and will be the strongest chess event (ever!). A total of 44 participants, including the Swiss 6 champion LcZero, the top rated Stockfish and exciting newcomers Ceres, PlentyChess, and Patricia, will battle it out for the title in this record breaking Swiss.

Swiss 7 will be daily live on the official website of TCEC, in video format on TCEC Twitch TV, a broadcast on our favorite game server Lichess, and news reports on Chessdom.

Breaking the ELO record

ELO records depend on multiple factors. They include, but are not limited to, the formula used, the field, the sample size (number of games), and even the configuration (when talking about computer chess). Luckily, TCEC has the precise measurements to detect rating jumps in chess software. And the jump in Swiss 7 is huge! With high confidence we can claim that this is by far the strongest chess event in history.

Putting the record into perspective and “human terms”: imagine a classical tournament at a time where Carlsen and Kasparov close to 3000 ELO and Caruana, Aronian, Firouzja, Topalov, Anand, So, Nakamura, Kramnik, and Giri are all above and around 2900 ELO. Now this is what we are viewing in computer chess at TCEC Swiss 7!

Here is a comparison of the top 10 participants in Swiss 1 and Swiss 7 events. The difference of rating jumps is even bigger among the rest of the participants.

Breaking the participation record

The state of computer chess – thriving! The sector has experienced a recent !boom with the development of new technology and ideas. For the first time ever TCEC had more applications than the event can accommodate. As a result, various engines had to undergo qualifications tournaments, raising the average ELO of the field even more.

TCEC Swiss 7 seeding and participants

This is the full participants list of TCEC Swiss 7 according to seeding. The seed number is determined according to the season rules, starting with the previous champion of TCEC – LcZero

1. LCZero (defending champion of TCEC Swiss 6)
2. Stockfish (Season 26 league champion)
3. Berserk
4. KomodoDragon
5. Ethereal
6. Seer
7. Ceres
8. RubiChess
9. Stoofvlees
10. Caissa
11. Obsidian
12. rofChade
13. Viridithas
14. Igel
15. Revenge
16. Arasan
17. Clover
18. Lizard
19. Velvet
20. PlentyChess
21. Minic
22. Uralochka
23. Stormphrax
24. Ginkgo
25. DeepSjeng
26. BlackMarlin
27. Marvin
28. Altair
29. Equisetum
30. Booot
31. Weiss
32. Tucano
33. Halogen
34. Renegade
35. akimbo
36. Texel
37. Winter
38. Devre
39. ChessFighter
40. Patricia
41. Princhess
42. ice4
43. 4ku
44. Heimdall

Which record is not broken?

TCEC Swiss 7 is part of the 27th edition of TCEC. It is one of the four titles that the engines can compete for, namely the Swiss, FRC, Leagues, and Cup titles. Inherently, every season is stronger than the previous one. Engine chess is booming, authors are more active than ever, technology is developing, ideas are growing. This season and the TCEC Swiss 7 event seems to break all records, but one. It is the TCEC Leagues S11 record, where the Premier Division average rating grew by 150+ ELO points. The average ELO of top 10 engines then was 3351 points, a full 150 points jump.

Another record in sight is the audience. The most viewed season of TCEC was 2 million people audience. Back then it was fueled by Carlsen’s World Chess Championships success. Since then TCEC seasons have crossed 1 million viewers several times, but has never been close to this record.

TCEC 4k IV standings

1. Ice4 43,5/48
2. 4ku 40,0
3. Stro4k 36,0
4. Molly 20,5
5. M4sseur 19,5
6. Moonfish 5,0
7. Pygone 3,5

Next: the 4k IV mini final, watch live at Visit the TCEC 4k live page / Follow TCEC on Twitch

From time to time TCEC treats us with side events. Currently ongoing is one of the most amazing such bonuses – the TCEC 4k competition. The key: all participating engines are limited to to a maximum of 4KiB (4096 bytes)

In modern age it is a challenge to find applications of this size. At TCEC, however, the competitors not only meet the requirement, but play above GM level. The top two engines – Ice4 and 4ku – even participate successfully in the regular season, boasting ELO of over 3000.

Visit the TCEC 4k live page / Follow the games on Twitch

The match started with Leela taking the lead 2,5-0,5 after winning a game with white in round 1 in Paulsen attack and a second game in round 3 in Sicilian Najdorf. Stockfish reacted, bringing the score level by game 10 and taking a two point lead afterwards. And just when everyone thought Stockfish might be on the way to an easy title, Leela Chess Zero showed that the games from the first rounds were not a coincidence and returned a point.

Replay the decisive games

Lc0 – Stockfish (Center game, Paulsen attack) 1-0
Lc0 – Stockfish (Sicilian Najdorf) 1-0
Stockfish – Lc0 (Sicilian Najdorf) 1-0
Stockfish – Lc0 (King’s Indian, Averbakh system) 1-0
Stockfish – Lc0 (Pirc, Austrian attack) 1-0
Stockfish – Lc0 (French, Tarrash) 1-0
Lc0 – Stockfish (King’s Indian, 5.Be2)

Currently the score is Stockfish 10,5 – Leela Chess Zero 9,5, with a game 22 (a Budapest) going on. Games are going on 24/7 and you can follow then with detailed analysis at the official website here

Unfortunately for the romantics, the first round served up no “Cupsets”. Stoofvlees struggled a little against Marvin, which is always tense because Stoof has a history of “oofing” and losing, but it didn’t happen this time. Round 2 saw a mini-cupset as 14th seed Koivisto put 6th seed Slowchess to the sword, but this wasn’t too surprising either because Koivisto had suffered from a major bug in the previous season and was seeded unusually low as a result. The rest of the matches all proceeded like clockwork as the higher-seeded engines bested their opponents.
But if people were beginning to lose hope, the semifinals brought a real surprise when Stockfish lost an opening to Komodo Dragon. Komodo Dragon is a strong engine, but it has not won an opening against Stockfish for a long time, with the -19 +0 =31 minimatch loss to Stockfish in the Season 22 superfinal an especially bad memory. There was no fairytale comeback this time as Stockfish recovered to win 6-4, but the dropped minimatch did make it seem like the otherwise-untouchable Stockfish can be beaten. Was it a blip, or was it a sign that possibly, just possibly, that Stockfish would fail to win the Cup? In the other semifinal, Lc0 dusted Ethereal away 5-3 to take up the reins as Stockfish’s final opponent. Things started out badly for Leela as Stockfish drew first blood in games 3/4, winning a French Defense, no less. Historically it had generally been Leela who was better at that opening. But in game 7, Leela demonstrated superior positional understanding after Stockfish failed to sense the danger of allowing its Bishop to be forced into a corner.

(Caption) A critical moment in game 7. White’s position is clearly pleasant, and Leela’s +0.97 evaluation is very high by Leela’s standards, but the game isn’t over by any means. Stockfish chose 19…O-O, allowing 20. h6 Bh8. Black’s Bishop is in trouble and the continuation 21. f3 f5 22. g5 made it even more miserable. Although superficially it looks like Black can open a diagonal for the Bishop with …c5, White only has to respond with c3 to keep the Bishop trapped.
This sparkling win tied the match. A tense series of draws later, Leela won game pair 15/16 to win the match. The mighty Stockfish was beaten for the first time in what feels like forever, and Leela won its first title since Season 17.
Next up was the Swiss. This format rewards engines that can reliably beat lower-rated engines, and Leela’s developers submitted a version with dynamic contempt for the first time. This feature is intended to make Leela play for wins against lower-rated opponents, and had already been used to great success by Stockfish (back when Stockfish used handcrafted eval) and Komodo Dragon. For the lower-rated engines, the Swiss took special significance because it affected placing in the Leagues – 22 spots in the Leagues were decided by placing from the previous season, with four more spots going to the top-placing non-qualified engines from the Swiss. 
I could highlight a lot of things, but one thing overshadows them all. In the very first round, Cup winner Leela produced the following brilliancy against QL engine Cheese:

Something had clearly gone wrong with Leela’s dynamic contempt. Needless to say Cheese won easily from this position, and although Leela also won with White to tie the match, it was already behind its rivals – both Stockfish and Komodo Dragon won their opening matches against strong opposition. When the same thing happened to Leela again in round 2 (this time against Drofa), it looked briefly like Leela might not finish in the top 3. Fortunately for Leela fans, the “Leelabug” doesn’t affect Leela’s play against strong engines, and the Swiss format even meant that Leela faced weaker opponents for the next few rounds. Leela steadily fought its way up the table to eventually finish 3rd. For the other members of the Big Three, Stockfish continued to show some weakness when it lost games – but not matches – to Revenge and Scorpio, although it did win the head-to-head against Komodo Dragon. That one point swing, combined with Leela’s dropped half-points against Drofa and Cheese, turned out to be just enough to give Stockfish the tournament victory – its final score of 15.5/22 was exactly one point ahead of Komodo Dragon and Leela on 14.5/22.

For other competitors, the qualifiers to League 2 turned out to be Velvet, Expositor, Wasp, and Uralochka. Velvet played some strong chess, including a 2-0 double kill over Expositor, to finish the highest of the four. Meanwhile, Expositor benefited heavily from the Swiss format when it was paired against the bottom three finishers, scoring 5/6. Its 11.5/22 score actually placed it above League 1 engine Berserk, who had the misfortune of being matched against all of the Big Three. The two 4k engines (so-called because their size is limited to 4 kilobytes) ice4k and 4ku finished solidly last, well off the pace. The 4k handicap is a tough one. 4ku’s developer even deleted the “print eval” line to save space, making 4ku games rather mysterious to watch.

Next up is League 2, where the four qualifiers will face several seasoned competitors, the biggest names being former Premier Division engines Igel and Fritz. Unsurprisingly, both engines are favorites for promotion. Testing is currently under way, and the league is slated to start right after. All games will be played live at https://tcec-chess.com/

Article by Low

But just as Leela seemed down, she turned on god mode. Leela won 10 of the next 22 openings before Komodo finally won the 35th opening. One “busted” opening and a traded win later, Leela won another five openings to round out the superfinal. The subfinal did turn into a rout, but in Leela’s favor. The dramatic comeback from -3 to +11 speaks to the dangers of small sample size as well. Even the full 100 games is not many compared to the tens of thousands of games that engine developers regularly play when testing their patches!

There was still the “bookmaker’s subfinal”, between GM Matthew Sadler and Jeroen Noomen. The two authored the superfinal book, with Jeroen running out the victor in that match by one decisive opening. This bookmaker’s subfinal also turned out very close, and in fact uncannily resembled the superfinal: the two bookmakers remained neck and neck through most of the subfinal, involved a dramatic game 100 (Leela failed to convert a position that Komodo evaluated at +4), and ended with the same score – Jeroen wins by one opening. This time though, GM Sadler has the consolation prize of getting more wins for one engine. Jeroen only scored two Komodo wins, but GM Sadler got five. Of course, this also means that Jeroen got 11 Lc0 wins, while GM Sadler only got seven.

Next up are several more bonus matches: there are some very long time control games (6+ hours per side per game), some DFRC games (Double Fischer Random Chess, where both sides start with a different-but-valid FRC position), and exhibition games by the top engines against the best of the chasing pack. The pause gives all competitors a time to catch their breath before Season 24. Find all bonus games at https://tcec-chess.com/. 

Article by Low