Dolphin 2.1 - new open-source UCI engine in C# with a self-trained NNUE

Discussion of anything and everything relating to chess playing software and machines.

Moderator: Ras

User avatar
AdminX
Posts: 6411
Joined: Mon Mar 13, 2006 2:34 pm
Location: Acworth, GA

Re: Dolphin 2.1 - new open-source UCI engine in C# with a self-trained NNUE

Post by AdminX »

Damir wrote: ↑Sun Oct 18, 2026 9:16 pm Hi Ted

Have you tried Deep Shredder 14. It has been released a little over a month ago by SMK on its official website.
Hi,

Not yet, but I plan on buying it. I think I qualify for a discount.
"Good decisions come from experience, and experience comes from bad decisions."
__________________________________________________________________
Ted Summers
User avatar
Graham Banks
Posts: 46643
Joined: Sun Feb 26, 2006 10:52 am
Location: Auckland, NZ

Re: Dolphin 2.1 - new open-source UCI engine in C# with a self-trained NNUE

Post by Graham Banks »

nguyenviethuy wrote: ↑Tue Oct 13, 2026 1:34 pm
Gabor Szots wrote: ↑Tue Oct 13, 2026 9:06 am I can see that the link we give for version 1.0 is no longer valid. Can that engine still be downloaded from some place of yours?
The old link is no longer in use; I am using only this new link.
I hope you will use this new version to replace the old Dolphin.
I post all the latest versions here: https://github.com/aiagentvn/dolphin/releases
When I try to create the engine in ChessGUI, it doesn't load properly (no author name, no parameters).
gbanksnz at gmail.com
nguyenviethuy
Posts: 7
Joined: Sun Oct 11, 2026 6:24 am
Full name: NGUYEN VIET HUY

Re: Dolphin 2.1 - new open-source UCI engine in C# with a self-trained NNUE

Post by nguyenviethuy »

Graham Banks wrote: ↑Sat Oct 24, 2026 12:36 am
nguyenviethuy wrote: ↑Tue Oct 13, 2026 1:34 pm
Gabor Szots wrote: ↑Tue Oct 13, 2026 9:06 am I can see that the link we give for version 1.0 is no longer valid. Can that engine still be downloaded from some place of yours?
The old link is no longer in use; I am using only this new link.
I hope you will use this new version to replace the old Dolphin.
I post all the latest versions here: https://github.com/aiagentvn/dolphin/releases
When I try to create the engine in ChessGUI, it doesn't load properly (no author name, no parameters).
Hi Graham,

Thank you for the report. I found the cause: when a GUI starts the engine without a console window, Dolphin 2.2.0 crashed at startup before it could answer "uci", so the GUI got no name, author or options.

This is fixed in Dolphin 2.2.1: https://github.com/aiagentvn/dolphin/re ... tag/v2.2.1

Playing strength is unchanged (same search and network, same bench 1627584). Please remove the old entry in ChessGUI and add 2.2.1 again. I would be grateful if you could tell me whether it loads correctly now.
nguyenviethuy
Posts: 7
Joined: Sun Oct 11, 2026 6:24 am
Full name: NGUYEN VIET HUY

Re: Dolphin 2.1 - new open-source UCI engine in C# with a self-trained NNUE

Post by nguyenviethuy »

Hi all,

Dolphin 2.4 is out. It is my open-source UCI engine written in C#, with an NNUE network trained only on Dolphin's own self-play games.

Download (Windows x64): https://github.com/aiagentvn/dolphin/re ... tag/v2.4.0
Source (MIT): https://github.com/aiagentvn/dolphin

First, thank you to the CCRL team: Dolphin 2.2 is now rated 3448 in CCRL Blitz. Thanks also to Graham Banks for reporting that the engine did not load in ChessGUI. It crashed at startup when a GUI started it without a console window. This was fixed in 2.2.1.

What's new since 2.2

2.3: search update (same network). Five changes, each kept only after a self-play SPRT at a fixed 40,000 nodes per move:
  • Correction history from five sources: pawn structure, each side's other pieces, and the last two move pairs.
  • No reverse futility pruning at TT-PV nodes, or when the hash move is quiet.
  • Evaluation scaled down as material comes off the board.
  • Fail-high scores pulled toward beta at non-PV nodes.
  • Continuation-history penalty for the opponent's previous move when it is refuted at once.
Against 2.2.1 in cutechess-cli at 10+0.1: 145-56-299, +62.5 ± 19.1 Elo (500 games). This is self-play, so the gain against other engines will most likely be smaller.

2.4: Syzygy tablebases, through Fathom (MIT). New UCI options SyzygyPath and SyzygyProbeDepth.
  • Root moves are filtered with DTZ, and WDL is probed in the search after captures and pawn moves. Info lines show tbhits.
  • fathom.dll is included in both packages. Without it, or without tables, the engine runs normally.
  • Without tablebases, 2.4 plays exactly like 2.3 (same bench: 1717319).
Testing with the 3-4-5 piece tables:
  • Same build, tablebases on vs off, 10+0.1: +11.8 ± 11.9 Elo (1000 games).
  • In 720 games that reached 5 pieces or fewer, the tablebase side never scored below the theoretical result.
  • Against the 2.3 package the cutechess results were slightly negative (-19.7 ± 23.9 in 300 games), within the error margin. The full numbers, good and bad, are in the release notes.
I only have 5-piece tables here, so 6-piece setups have not been tested. Feedback from anyone using them is very welcome.

Verify: "Dolphin2.exe bench" should print 1717319 nodes, and "perft 5" should print 4865609.

Bug reports and test results are always welcome, here or on GitHub.

Nguyen Viet Huy (Vietnam)
nguyenviethuy
Posts: 7
Joined: Sun Oct 11, 2026 6:24 am
Full name: NGUYEN VIET HUY

Re: Dolphin 2.1 - new open-source UCI engine in C# with a self-trained NNUE

Post by nguyenviethuy »

Hi all,

A quick follow-up: Dolphin 2.5 is out, a few hours after 2.4.

Download (Windows x64): https://github.com/aiagentvn/dolphin/re ... tag/v2.5.0

What's new: a new network (net6). Search and Syzygy support are the same as in 2.4.
  • Trained on about 54 million positions from six generations of Dolphin's own self-play games. The sixth generation, 9.7 million positions, was played with the 2.4 search.
  • Same size as before (768->384)x2, and no data or networks from other engines.
Testing (self-play, so the gain against other engines will most likely be smaller):
  • New vs old network in the same engine: +16.2 ± 11.7 Elo at 20,000 nodes per move (2000 games), and +18.8 ± 12.9 Elo at 10+0.1 (1000 games).
  • Against the 2.4 release package in cutechess-cli at 10+0.1, both with Syzygy 3-4-5: 119-91-290, +19.5 ± 19.7 Elo (500 games).
  • 720 games in total, with no time losses, crashes or illegal moves.
Verify: "Dolphin2.exe bench" should print 1551453 nodes, and "perft 5" should print 4865609.

If you are testing 2.4, please use 2.5 instead: it has the same features and a stronger network.

Nguyen Viet Huy (Vietnam)