Luna CE 4.1.1
https://github.com/Spunc595/Luna-Chess- ... tag/v4.1.1
Small corrective release of 4.1.0. The Windows exe is now a baseline
x86-64 build (the 4.1.0 exe required AVX2/BMI and could crash on older
CPUs). Also a Move Overhead option and cleaner UCI startup output.
Search and evaluation are unchanged from 4.1.0.
UCI engine written in Rust, NNUE network embedded. Binaries: Windows
x86-64, Linux aarch64, Android arm64. Defaults: Hash 256, Threads 1.
Estimated strength: about the same as 4.1.0. For reference, Luna CE
4.0.0 is rated 3334 on CCRL Blitz.
New engine releases 2026 H2
Moderator: Ras
-
Spunc595
- Posts: 44
- Joined: Mon Jul 06, 2026 12:15 am
- Full name: Daniele Marpino
-
CMCanavessi
- Posts: 1151
- Joined: Thu Dec 28, 2017 4:06 pm
- Location: Argentina
Re: New engine releases 2026 H2
Facón 1.8 - Cabo
https://github.com/CMCanavessi/facon/releases/tag/v1.8
Approximately +300 Elo over 1.7 (Ordo ~3230, Gauntlet 5 n=1040). Direct self-play vs 1.7: +346.7 Elo (n=10000, +/- 7.3) at 10+0.1.
Where 1.7 split its gain between search and evaluation, 1.8 is the search release: the evaluation is 1.7's, unchanged. It fixes every bug a full audit of 1.7 found, makes the engine about 70% faster before touching the search, then rebuilds the search -- principal variation search, a transposition table on cache-line buckets that quiescence uses too, correction history, and reworked null-move pruning, late move reductions and futility pruning -- and adds parallel search (Lazy SMP) and Syzygy endgame tablebases.
Development followed 1.7's discipline: every change was added alone, measured in self-play against the previous accepted build, and kept only if it measured well. Three long checkpoint matches against 1.7 kept the accumulated total honest along the way: +135.3 (n=10000), +189.0 (n=3220) and +219.6 (n=8120).
Reliability
Every bug from the audit of 1.7 fixed: go followed at once by stop no longer hangs; position validates its FEN and moves and leaves the board untouched when they are invalid, saying why; go movetime, go nodes and go infinite do what they say; clocks of any size, and at zero or below, are handled. None of these triggered in rating conditions.
Hash gets exactly the size requested, up to 64 TiB, without the cap to available RAM that made a re-sent value oscillate.
Speed
One evaluation walk computes the evaluation, the trace decomposition and the eval breakdown, with every piece's attacks computed once (+17% NPS, about +41 Elo); material and piece-square sums are kept up to date as moves are made (+5% NPS); moves are generated in stages, only as the search reaches them (+42% NPS, about +76 Elo). About 70% more nodes per second than 1.7 before any change to the search.
Search
Exact en passant and evasions: the en passant square is set only when a legal capture exists, so positions hash one way and repetitions are detected (about +10 Elo); quiescence searches every evasion when in check (about +12 Elo).
Transposition table: four entries per 64-byte cache line, validated without locks, any size, huge pages on Linux (about +11 Elo); quiescence reads and writes it, with fail-soft scores (about +32 Elo).
Principal variation search, with reverse futility, null-move pruning, razoring, ProbCut and late move pruning kept off the principal variation and reductions at expected cut nodes. About +32 Elo -- both forms tried in 1.7 had measured negative.
Pruning and reductions reworked: null-move pruning gated on the evaluation, with an adaptive reduction (about +15 Elo); late move reductions from a logarithmic table with node, history and improvement adjustments (about +18 Elo); futility pruning to depth 10 (about +14 Elo); late move reductions in check, the aspiration window after a score of 0, and an interrupted iteration's better move played (about +20 Elo).
Correction history: seven tables learn where the static evaluation misses -- by pawn structure, minor pieces, major pieces, each side's other pieces, and the last two moves -- and correct it in the main search and in quiescence. About +85 Elo, the largest single gain of the release.
History across searches: the butterfly history is halved at each new search instead of cleared, with a larger step for a cutoff. About +26 Elo.
Parallel search
Lazy SMP, with a new Threads option up to 1024: helper threads share the transposition table and the correction history, and the move played is chosen by a vote of the threads. Four threads search almost four times as many nodes per second as one. The rating figures above are single-threaded.
Endgame tablebases
Syzygy tablebases, read through Pyrrhic (Andrew Grant, MIT license), with the options SyzygyPath, SyzygyProbeDepth and SyzygyProbeLimit: win/draw/loss probes in the search (about +16 Elo), and the root's moves ranked by distance to zeroing, so that wins close to the 50-move limit are converted -- 80 of 80 in a test where the probes alone converted none.
Features attempted and not shipped
SEE pruning in the main search: neutral over 8,160 games, and 8% slower.
A correction to the singular-extension verification: -3.3 Elo over 9,308 games.
Check extension near the horizon only: neutral; the extension stays at every depth.
Screened and dropped: capture history, history pruning, two-ply continuation history, multicut, mate distance pruning, quiet checks in quiescence, double and negative extensions.
Checksums (SHA1)
63d27d1f856509f663c1b6f75479f86e1568bd3a facon-1.8
3d5f6d994ea2da6b87c97160b4f3be924066ed71 facon-1.8.exe
https://github.com/CMCanavessi/facon
https://github.com/CMCanavessi/facon/releases/tag/v1.8
Approximately +300 Elo over 1.7 (Ordo ~3230, Gauntlet 5 n=1040). Direct self-play vs 1.7: +346.7 Elo (n=10000, +/- 7.3) at 10+0.1.
Where 1.7 split its gain between search and evaluation, 1.8 is the search release: the evaluation is 1.7's, unchanged. It fixes every bug a full audit of 1.7 found, makes the engine about 70% faster before touching the search, then rebuilds the search -- principal variation search, a transposition table on cache-line buckets that quiescence uses too, correction history, and reworked null-move pruning, late move reductions and futility pruning -- and adds parallel search (Lazy SMP) and Syzygy endgame tablebases.
Development followed 1.7's discipline: every change was added alone, measured in self-play against the previous accepted build, and kept only if it measured well. Three long checkpoint matches against 1.7 kept the accumulated total honest along the way: +135.3 (n=10000), +189.0 (n=3220) and +219.6 (n=8120).
Reliability
Every bug from the audit of 1.7 fixed: go followed at once by stop no longer hangs; position validates its FEN and moves and leaves the board untouched when they are invalid, saying why; go movetime, go nodes and go infinite do what they say; clocks of any size, and at zero or below, are handled. None of these triggered in rating conditions.
Hash gets exactly the size requested, up to 64 TiB, without the cap to available RAM that made a re-sent value oscillate.
Speed
One evaluation walk computes the evaluation, the trace decomposition and the eval breakdown, with every piece's attacks computed once (+17% NPS, about +41 Elo); material and piece-square sums are kept up to date as moves are made (+5% NPS); moves are generated in stages, only as the search reaches them (+42% NPS, about +76 Elo). About 70% more nodes per second than 1.7 before any change to the search.
Search
Exact en passant and evasions: the en passant square is set only when a legal capture exists, so positions hash one way and repetitions are detected (about +10 Elo); quiescence searches every evasion when in check (about +12 Elo).
Transposition table: four entries per 64-byte cache line, validated without locks, any size, huge pages on Linux (about +11 Elo); quiescence reads and writes it, with fail-soft scores (about +32 Elo).
Principal variation search, with reverse futility, null-move pruning, razoring, ProbCut and late move pruning kept off the principal variation and reductions at expected cut nodes. About +32 Elo -- both forms tried in 1.7 had measured negative.
Pruning and reductions reworked: null-move pruning gated on the evaluation, with an adaptive reduction (about +15 Elo); late move reductions from a logarithmic table with node, history and improvement adjustments (about +18 Elo); futility pruning to depth 10 (about +14 Elo); late move reductions in check, the aspiration window after a score of 0, and an interrupted iteration's better move played (about +20 Elo).
Correction history: seven tables learn where the static evaluation misses -- by pawn structure, minor pieces, major pieces, each side's other pieces, and the last two moves -- and correct it in the main search and in quiescence. About +85 Elo, the largest single gain of the release.
History across searches: the butterfly history is halved at each new search instead of cleared, with a larger step for a cutoff. About +26 Elo.
Parallel search
Lazy SMP, with a new Threads option up to 1024: helper threads share the transposition table and the correction history, and the move played is chosen by a vote of the threads. Four threads search almost four times as many nodes per second as one. The rating figures above are single-threaded.
Endgame tablebases
Syzygy tablebases, read through Pyrrhic (Andrew Grant, MIT license), with the options SyzygyPath, SyzygyProbeDepth and SyzygyProbeLimit: win/draw/loss probes in the search (about +16 Elo), and the root's moves ranked by distance to zeroing, so that wins close to the 50-move limit are converted -- 80 of 80 in a test where the probes alone converted none.
Features attempted and not shipped
SEE pruning in the main search: neutral over 8,160 games, and 8% slower.
A correction to the singular-extension verification: -3.3 Elo over 9,308 games.
Check extension near the horizon only: neutral; the extension stays at every depth.
Screened and dropped: capture history, history pruning, two-ply continuation history, multicut, mate distance pruning, quiet checks in quiescence, double and negative extensions.
Checksums (SHA1)
63d27d1f856509f663c1b6f75479f86e1568bd3a facon-1.8
3d5f6d994ea2da6b87c97160b4f3be924066ed71 facon-1.8.exe
https://github.com/CMCanavessi/facon
My chess engine Facón: https://github.com/CMCanavessi/facon
-
LFM
- Posts: 1
- Joined: Fri Oct 30, 2026 5:35 pm
- Full name: Louis Simonet
Re: New engine releases 2026 H2
Hey everyone,
Would like to announce the release of my custom-built engine Liora v11.0 F, written in C
Estimated strength: 2614 ± 28.9 in pure HCE (calculated from 1k games against Stockfish 19 with UCI_Elo set at 2350)
NNUE is supported but not yet trained (engine defaults to HCE). Pre-trained networks will be available on GitHub in the future.
Supports standard UCI commands
Liora is single threaded and does not support Syzygy tablebase.
All assets (including Windows and Linux x64 binaries and rating details) are available at https://github.com/LFM57/liora-chess-en ... ses/latest
Written by me, LFM (France)
I'm opened to any question!
LFM
Would like to announce the release of my custom-built engine Liora v11.0 F, written in C
Estimated strength: 2614 ± 28.9 in pure HCE (calculated from 1k games against Stockfish 19 with UCI_Elo set at 2350)
NNUE is supported but not yet trained (engine defaults to HCE). Pre-trained networks will be available on GitHub in the future.
Supports standard UCI commands
Liora is single threaded and does not support Syzygy tablebase.
All assets (including Windows and Linux x64 binaries and rating details) are available at https://github.com/LFM57/liora-chess-en ... ses/latest
Written by me, LFM (France)
I'm opened to any question!
LFM