New engine releases & news H1 2026

[pkrisz]

Renegade 1.3.0 - https://github.com/pkrisz99/Renegade/re ... tag/v1.3.0

Self-play against version 1.2.0 with an unbalanced book:

Code: Select all

Elo   | 91.56 +- 3.40 (95%)
Conf  | 50.0+0.50s Threads=1 Hash=128MB
Games | N: 10000 W: 3775 L: 1199 D: 5026
Penta | [0, 319, 1923, 2621, 137]

[NathanDrake]

Triumviratus 4.1 — release + a question about CCRL

Hi all,

A while back I posted Triumviratus 4.0 here and some of you were kind enough to give it a spin — thanks for the feedback, it really helped.

Triumviratus 4.1 is out: https://github.com/Tors3/Triumviratus. Source + Windows AVX2 + AVX-512 PGO builds inside.

What's new in 4.1

• Large SPSA co-tune (~10k iterations on a 176-vCPU instance, ~18 parameters across pruning margins, history weights, LMR + SF-style depth-pruning block).
• SPRT vs 4.0, same NNUE both sides: +35.0 ±17.9 Elo @ 20+0.2 (LOS ~100%, fastchess + UHO_2024 +080/+099). Bench 2601747.
• No new search features vs 4.0, same architecture — just better-tuned constants.

Two questions if anyone's around:

• How does one get listed on CCRL? I'm a hobbyist and a bit fuzzy on the process — is there a submission form somewhere, or do you just put a clean binary + readme out and hope a tester picks it up? Any pointers from someone who's been through it would be much appreciated.
• Any feedback welcome — bench, odd positions, NPS on your hardware, whatever you've got.

Looking ahead

• A 5.0 with SFNNv10 + Full_Threats (the newer SF-style arch) is in progress.
• Training on publicly available datasets only for now — Vondele's SF-gen packs from Kaggle + Leela test80 binpacks. No compute for serious self-play yet, that's months away.
• The thing I'd love to get tested now is 4.1.

Cheers

[Graham Banks]

Triumviratus 4.1 — release + a question about CCRL
How does one get listed on CCRL? I'm a hobbyist and a bit fuzzy on the process — is there a submission form somewhere, or do you just put a clean binary + readme out and hope a tester picks it up? Any pointers from someone who's been through it would be much appreciated.

Evaluation: NNUE (Stockfish HalfKAv2_hm architecture) with dual nets (big/small), a single-board bridge feeding the network from the engine's native bitboards, a finny-table accumulator-refresh cache, a per-thread static-eval cache, a 16-bit static eval cached in the transposition table, and learned correction history (pawn + minor + major material keys).

I seem to recall that you were going to develop your own net. That is why I've held off testing it for now.

[NathanDrake]

Triumviratus 4.1 — release + a question about CCRL
How does one get listed on CCRL? I'm a hobbyist and a bit fuzzy on the process — is there a submission form somewhere, or do you just put a clean binary + readme out and hope a tester picks it up? Any pointers from someone who's been through it would be much appreciated.

Evaluation: NNUE (Stockfish HalfKAv2_hm architecture) with dual nets (big/small), a single-board bridge feeding the network from the engine's native bitboards, a finny-table accumulator-refresh cache, a per-thread static-eval cache, a 16-bit static eval cached in the transposition table, and learned correction history (pawn + minor + major material keys).

I seem to recall that you were going to develop your own net. That is why I've held off testing it for now.

Hi Graham,

Yes, own net is still the plan. Training is running now (SFNNv10 + Full_Threats, public datasets for the moment) but I'm a student, so realistically a few weeks to a month before there's a proper own-net release.

IIf you happen to have a bit of time for 4.1 in the meantime I'd really appreciate it, making it onto CCRL would honestly mean a lot to me personally. Totally understand if you'd rather wait for the full thing. Will ping you once 5.0 is up.

Thanks.

[Gabor Szots]

Triumviratus 4.1 — release + a question about CCRL
How does one get listed on CCRL? I'm a hobbyist and a bit fuzzy on the process — is there a submission form somewhere, or do you just put a clean binary + readme out and hope a tester picks it up? Any pointers from someone who's been through it would be much appreciated.

Evaluation: NNUE (Stockfish HalfKAv2_hm architecture) with dual nets (big/small), a single-board bridge feeding the network from the engine's native bitboards, a finny-table accumulator-refresh cache, a per-thread static-eval cache, a 16-bit static eval cached in the transposition table, and learned correction history (pawn + minor + major material keys).

I seem to recall that you were going to develop your own net. That is why I've held off testing it for now.

This is exactly why I haven't tested it yet.

[NathanDrake]

Triumviratus 4.1 — release + a question about CCRL
How does one get listed on CCRL? I'm a hobbyist and a bit fuzzy on the process — is there a submission form somewhere, or do you just put a clean binary + readme out and hope a tester picks it up? Any pointers from someone who's been through it would be much appreciated.

Evaluation: NNUE (Stockfish HalfKAv2_hm architecture) with dual nets (big/small), a single-board bridge feeding the network from the engine's native bitboards, a finny-table accumulator-refresh cache, a per-thread static-eval cache, a 16-bit static eval cached in the transposition table, and learned correction history (pawn + minor + major material keys).

I seem to recall that you were going to develop your own net. That is why I've held off testing it for now.

This is exactly why I haven't tested it yet.

Hi Gabor, Graham,

Thanks for letting me know. Do I need my own network to be tested for CCRL? If yes, I'll release a 4.2 with my older own-trained net (rubicon). Slightly weaker, but fully my own.

Thanks

[Graham Banks]

Triumviratus 4.1 — release + a question about CCRL
How does one get listed on CCRL? I'm a hobbyist and a bit fuzzy on the process — is there a submission form somewhere, or do you just put a clean binary + readme out and hope a tester picks it up? Any pointers from someone who's been through it would be much appreciated.

Evaluation: NNUE (Stockfish HalfKAv2_hm architecture) with dual nets (big/small), a single-board bridge feeding the network from the engine's native bitboards, a finny-table accumulator-refresh cache, a per-thread static-eval cache, a 16-bit static eval cached in the transposition table, and learned correction history (pawn + minor + major material keys).

I seem to recall that you were going to develop your own net. That is why I've held off testing it for now.

This is exactly why I haven't tested it yet.

Hi Gabor, Graham,

Thanks for letting me know. Do I need my own network to be tested for CCRL? If yes, I'll release a 4.2 with my older own-trained net (rubicon). Slightly weaker, but fully my own.

Thanks

Either that or we can wait.

[NathanDrake]

I would prefer to go with the current 4.1 for the CCRL rating.

In the next weeks/months I will come back with my own net and a big update of the engine, but for now I would prefer for you to test the current version, if possible.

Thank you, let me know if you need any details.

[Graham Banks]

I would prefer to go with the current 4.1 for the CCRL rating.

In the next weeks/months I will come back with my own net and a big update of the engine, but for now I would prefer for you to test the current version, if possible.

Thank you, let me know if you need any details.

Only if you use your own net with this version (4.1).
Otherwise, we'll wait.

[NathanDrake]

I would prefer to go with the current 4.1 for the CCRL rating.

In the next weeks/months I will come back with my own net and a big update of the engine, but for now I would prefer for you to test the current version, if possible.

Thank you, let me know if you need any details.

Only if you use your own net with this version (4.1).
Otherwise, we'll wait.

Ok, I will add two more .zip files to the release with my own net. The engine uses a small net as well (also from SF). Do I need to remove that one too? Or is it sufficient to have my own net for the big net?