Gaia 4.1.0 — Back after 20 years

[j_romang]

Hi everyone,

I'm happy to announce the release of Gaia 4.1.0, available here:
https://github.com/jromang/gaiachess/releases/tag/4.1.0

Some of you may remember Gaia from the early 2000s — a small French engine written in C by David Rabel and myself between 2003 and 2006. After almost 20 years of silence, Gaia is back, completely rewritten from scratch in Rust.

What's in there:

NNUE evaluation (HalfKA with compact threat features, 13 king buckets, 8 output buckets, AVX-512)
PVS with aspiration windows, NMP, LMR, Singular Extensions, RFP/FP/LMP, SEE pruning, History Pruning, ProbCut, IIR
Full move ordering: killers, countermove, butterfly/capture/continuation/pawn history
Correction history (pawn, non-pawn, minor piece, continuation)
Lazy SMP, Syzygy tablebases, pondering, Multi-PV
Full UCI support (heavily untested )

Binaries are provided for windows x86-64 (AVX-512). The NNUE network is embedded in the binary.

About the source code: The code is not yet published — it's not mature enough to my taste. I want to clean it up properly before making it public. It will be open-sourced in the future.

I'd be curious to see where it lands on rating lists. Feedback and testing are very welcome!

Jean-François

[Graham Banks]

Welcome back.

Can you please give more information about the NNUE?
Also estimated strength.

The Windows binary won't run on my 5950x. Could you please provide an avx2/bmi2 compile?

[j_romang]

Thanks! Great to be back.

NNUE: HalfKA architecture (13 king buckets, 14,657 input features including compact threat/mobility features), 768 → SCReLU+pairwise → 32 → 64 → 1, 8 output buckets. AVX-512/AVX2/scalar compile-time dispatch, sparse L1, Finny table cache. Trained on only one generation for now

Strength: Honestly, I don't know — I haven't submitted to any rating lists yet, only did self-play SPRT testing. It should be significantly stronger than Gaia 3.5 from 2006 though.

AVX2 build: Sorry about that! The 5950X supports AVX2 but not AVX-512. I've added an AVX2/BMI2 build to the GitHub release, it should work fine on your Ryzen now.

[Graham Banks]

Thanks! Great to be back.

NNUE: HalfKA architecture (13 king buckets, 14,657 input features including compact threat/mobility features), 768 → SCReLU+pairwise → 32 → 64 → 1, 8 output buckets. AVX-512/AVX2/scalar compile-time dispatch, sparse L1, Finny table cache. Trained on only one generation for now

Strength: Honestly, I don't know — I haven't submitted to any rating lists yet, only did self-play SPRT testing. It should be significantly stronger than Gaia 3.5 from 2006 though.

AVX2 build: Sorry about that! The 5950X supports AVX2 but not AVX-512. I've added an AVX2/BMI2 build to the GitHub release, it should work fine on your Ryzen now.

Thanks for the extra exe.

You've trained your own NNUE and not borrowed one was my main concern.

[BrendanJNorman]

Hi everyone,

I'm happy to announce the release of Gaia 4.1.0, available here:
https://github.com/jromang/gaiachess/releases/tag/4.1.0

Some of you may remember Gaia from the early 2000s — a small French engine written in C by David Rabel and myself between 2003 and 2006. After almost 20 years of silence, Gaia is back, completely rewritten from scratch in Rust.

What's in there:

NNUE evaluation (HalfKA with compact threat features, 13 king buckets, 8 output buckets, AVX-512)
PVS with aspiration windows, NMP, LMR, Singular Extensions, RFP/FP/LMP, SEE pruning, History Pruning, ProbCut, IIR
Full move ordering: killers, countermove, butterfly/capture/continuation/pawn history
Correction history (pawn, non-pawn, minor piece, continuation)
Lazy SMP, Syzygy tablebases, pondering, Multi-PV
Full UCI support (heavily untested )

Binaries are provided for windows x86-64 (AVX-512). The NNUE network is embedded in the binary.

About the source code: The code is not yet published — it's not mature enough to my taste. I want to clean it up properly before making it public. It will be open-sourced in the future.

I'd be curious to see where it lands on rating lists. Feedback and testing are very welcome!

Jean-François

Legendary engine.

I remember Gaia 2.2 was a big favorite because it had this stodgy positional style like an old master haha.

Good for human opponents.

Would have been interesting to train a NNUE on this old Gaia's games and see what comes from it.

Rewritten is cool too though. Will have a look. Thanks and welcome back!

[j_romang]

Thanks, that really means a lot! I'm genuinely surprised and happy to see someone remembers Gaia from that era — we're talking over 20 years ago, that's ancient history in computer chess terms!
Gaia 2.2 and its "stodgy old master" style — that's a perfect description honestly, I love it.
That's actually a really cool idea about training a NNUE on old Gaia games — kind of distilling that old-school DNA into a modern framework. Who knows, maybe that's a fun experiment for the future!
In any case, I hope you enjoy this new version. It's a very different beast under the hood, but the spirit is still there. Let me know what you think if you give it a spin — any feedback is welcome!
Cheers,
Jean-François

[Gabor Szots]

Hi Jean-Francois,

That's incredible news! Gaia 3.5 is a regular participant of my tournaments, it's lucky that multiple changes in CPU and operation system have not made it unusable.
l am looking forward to testing the new one.

[Werner]

We found a bug: Engine cannot mate with KQ - K:
[pgn][Event "Div4 5s+600ms"]
[Site "isle of Rügen"]
[Date "2026.02.22"]
[Round "1"]
[White "Gaia 4.1.0"]
[Black "Sirius 9.0"]
[Result "1/2-1/2"]
[ECO "D35"]
[Opening "Damengambit"]
[Time "23:38:44"]
[Variation "Abtauschvariante, 5.Nf3"]
[TimeControl "5+0"]
[Termination "normal"]
[PlyCount "317"]
[WhiteType "program"]
[BlackType "program"]

1. d4 {+0.57/15 1} Nf6 2. c4 {+0.37/15 0} e6 3. Nf3 {+0.47/15 0} d5 4. cxd5
{+0.41/15 1} exd5 5. Nc3 {+0.57/15 0} Bd6 6. Bg5 {+0.79/14 0} c6 7. e3
{+0.83/14 0} h6 8. Bh4 {+1.01/14 0} Bf5 9. Qb3 {+1.27/15 0} b6 10. Ne5
{+0.98/12 0} O-O 11. Be2 {+0.64/14 1} Qe7 12. f4 {+1.10/12 0} b5 13. O-O
{+1.59/13 0} a5 14. Qd1 {+1.39/12 0} Qe6 15. Rc1 {+1.22/14 0} Nbd7 16. g4
{+2.85/15 0} Bh7 17. f5 {+4.63/15 0} Qe8 18. Nxc6 {+5.11/13 0} Qxe3+ 19.
Kg2 {+4.71/14 0} h5 20. g5 {+3.44/13 1} b4 21. Rf3 {+2.27/15 0} Ng4 22. Nb5
{+2.62/16 0} Bf4 23. Rxe3 {+1.89/17 0} Nxe3+ 24. Kf3 {+2.02/14 0} Nxd1 25.
Kxf4 {+2.34/17 0} Nxb2 26. Nc7 {+1.76/17 0} Rac8 27. Ne7+ {+4.98/15 0} Kh8
28. Nxc8 {+3.55/16 0} Rxc8 29. Rc2 {+3.62/17 0} Nb8 30. Bf2 {+4.33/14 1}
Na4 31. Bd3 {+4.52/15 0} Nb6 32. Rc5 {+4.91/14 0} N8d7 33. Ba6 {+5.57/17 0}
Nxc5 34. dxc5 {+5.19/16 0} Rxc7 35. cxb6 {+5.97/18 0} Re7 36. b7 {+6.06/17
0} Re4+ 37. Kf3 {+6.34/15 0} Re8 38. Ba7 {+6.35/16 0} Bxf5 39. Bb5
{+6.44/16 0} Rf8 40. b8=Q {+6.76/16 0} Rxb8 41. Bxb8 {+6.61/14 0} Bb1 42.
Bc7 {+7.12/15 0} Bxa2 43. Bxa5 {+7.58/15 0} b3 44. Bc3 {+6.75/18 0} Bb1 45.
Kf4 {+6.64/23 0} Kh7 46. Bb2 {+6.64/28 0} Kg6 47. Bc6 {+6.64/30 0} Be4 48.
Ba1 {+6.64/27 0} h4 49. Bd7 {+7.74/20 0} Bd3 50. Be8 {+7.27/24 0} Be2 51.
Bc6 {+7.65/22 0} Bc4 52. Bd4 {+7.58/22 0} Kh7 53. Ba4 {+7.57/21 0} Kg8 54.
Bb2 {+7.59/21 0} d4 55. Bxd4 {+8.10/19 0} f6 56. gxf6 {+8.56/18 0} gxf6 57.
Bxf6 {+9.93/19 0} Kf7 58. Bb2 {+11.40/16 0} Be6 59. Bc6 {+11.14/20 0} Kg6
60. Be8+ {+10.01/23 0} Kh6 61. Bb5 {+10.01/23 0} Bh3 62. Bc4 {+10.46/24 0}
Kg6 63. Bxb3 {+10.52/26 0} Kh6 64. Bc2 {+10.52/26 0} Kh5 65. Bf6 {+10.52/27
0} Bg2 66. Bd3 {+10.52/27 0} h3 67. Be5 {+10.07/24 0} Bd5 68. Bb5
{+11.58/28 0} Kg6 69. Be8+ {+11.58/33 0} Kh7 70. Kf5 {+11.58/34 0} Kg8 71.
Kf6 {+11.58/32 0} Kf8 72. Bd7 {+11.58/28 0} Bf3 73. Be6 {+11.91/32 0} Ke8
74. Bxh3 {+11.91/29 2} Be2 75. Bc7 {+19.43/22 1} Bd3 76. Be6 {+20.89/24 0}
Ba6 77. h4 {+20.89/20 0} Bf1 78. h5 {+M461/16 0} Bd3 79. Bg4 {+M458/17 0}
Be4 80. Bf5 {+M458/22 0} Bd5 81. Bf4 {+M458/19 1} Kd8 82. h6 Bg8 83. Bd6
{+M457/22 0} Ke8 84. Bd7+ {+M459/16 0} Kxd7 85. Kg7 {+M466/23 0} Bc4 86. h7
{+11.64/14 0} Kxd6 87. Kh6 {+M457/15 0} Bd3 88. h8=Q {+9.43/16 0} Kd5 89.
Qc3 {+9.77/21 0} Ke4 90. Kh5 {+M457/18 1} Be2+ 91. Kh4 {+212.39/17 0} Bf3
92. Qc6+ {+212.39/11 0} Ke3 93. Qc3+ {+166.49/16 0} Kf4 94. Kh3 Be4 95.
Qf6+ {+168.46/16 0} Ke3 96. Kg3 {+99.54/17 1} Bc2 97. Kg2 Kd3 98. Qb2
{+99.54/17 0} Bd1 99. Kg3 {+99.54/15 0} Bc2 100. Qb4 {+73.91/15 0} Ke3 101.
Qc3+ {+M471/19 0} Bd3 102. Qe5+ {+86.26/15 0} Be4 103. Qc5+ {+M470/18 0}
Kd3 104. Kf2 {+M454/20 0} Bh1 105. Qe3+ {+M453/21 0} Kc2 106. Qe6 {+M460/18
0} Kb2 107. Qe2+ {+M462/16 0} Ka3 108. Qc4 {+M458/20 0} Kb2 109. Ke1 Ka3
110. Kf2 {+M456/15 0} Bb7 111. Qc5+ {+M454/21 0} Ka4 112. Qa7+ {+M454/17 0}
Kb5 113. Qxb7+ {+M453/20 0} Kc5 114. Qa6 {+M448/18 0} Kd5 115. Qg6
{+M448/18 0} Ke5 116. Ke3 {+M448/17 1} Kd5 117. Kd2 Ke5 118. Ke3 {+M448/17
0} Kd5 119. Kd2 Ke5 120. Qa6 {+M448/16 0} Kd5 121. Qb5+ {+M448/17 0} Kd6
122. Qb8+ {+M448/15 0} Ke7 123. Qb7+ {+M448/16 0} Kd6 124. Kc1 Ke5 125. Qb6
{+M448/16 0} Kd5 126. Qg6 {+M448/16 1} Ke5 127. Kb1 Kf4 128. Qd6+ {+M448/16
0} Ke4 129. Kc2 {+M448/14 0} Ke3 130. Qd3+ {+M448/15 0} Kf4 131. Qh3
{+M448/17 0} Ke5 132. Kb1 Kd6 133. Qh6+ {+M448/16 0} Ke5 134. Ka1 Kf5 135.
Qb6 {+M448/17 0} Kg5 136. Kb2 {+M448/18 1} Kf5 137. Ka1 Ke5 138. Qa5+
{+M448/17 0} Kf6 139. Kb1 Kg6 140. Qd5 {+M448/18 0} Kf6 141. Qd6+ {+M448/18
0} Kg7 142. Ka1 Kh7 143. Qe7+ {+M448/19 0} Kh6 144. Qf6+ {+M455/17 0} Kh5
145. Qf5+ {+M448/20 0} Kh6 146. Kb1 {+M448/20 1} Kg7 147. Ka1 Kh8 148. Qd7
{+M472/20 0} Kg8 149. Kb1 Kf8 150. Qh7 {+M472/18 1} Ke8 151. Kb2 {+M13/24
0} Kf8 152. Qd7 {+262.76/17 0} Kg8 153. Kc3 {+M5/35 0} Kh8 154. Kd4 {+M4/51
0} Kg8 155. Ke5 {+M3/77 1} Kh8 156. Kf6 {+M2/77 1} Kg8 157. Ke5 Kh8 158.
Kf6 {+M2/78 1} Kg8 159. Ke5 {Stellungswiederholung} 1/2-1/2[/pgn]

[Sylwy]

We found a bug: Engine cannot mate with KQ - K:

3 men ? Syzygy enabled ? The engine can access it.

[Sylwy]

viewtopic.php?p=989770#p989770