Since you’re here...

We hope you will consider supporting us today. We need your support to continue to exist, because good entries are more and more work time. Every reader contribution, however big or small, is so valuable. Support "Chess Engines Diary" even a small amount– and it only takes a minute. Thank you.
============================== My email: jotes@go2.pl



Hypersion 3.0 - new version chess engine

on

Hypersion a free, open-source UCI chess engine.
Hypersion uses an NNUE network for evaluation (Stockfish 18 SFNNv10 architecture: HalfKAv2_hm features + FullThreats, big network 1024-d FT, small network 128-d FT) on top of an alpha-beta search with PVS, transposition table, aspiration windows, late-move reductions, singular extensions, ProbCut, futility/razoring/SEE pruning, and lazy-SMP infrastructure (single-threaded by default — multi-thread is currently unstable).
Author: RenCopp

Hypersion 3.0 what's new?

Patch notes (binaries updated 2026-05-11, second round)

Four additional commits land on top of the original v3.0:

  • Endgame mate-conversion fix (eb35855) — user-reported lichess
    game H05vgtVr where the engine drew K+R vs K by 50-move rule. In
    clearlyWinningEndgame positions (≤10 pieces and ≥rook material
    lead) the strength-limit caps now lift entirely so the engine can
    actually find a 16-ply mating plan. Validated: 245/252 = 97%
    conversion     across 21 cells (7 UCI_Elo levels × 3 movetimes).

  • Real UCI_AnalyseMode pruning relaxation (26fa62d) — when GUI
    sets UCI_AnalyseMode=true, the search now disables NMP, skips
    LMP/futility/SEE-margin pruning, and halves LMR aggression for
    thoroughness. Tactical-suite uplift at standard analysis conditions:

    SuiteBeforeAfterDelta
    WAC93.9%96.5%+2.6%
    mate-in-394.5%97.5%+3.0%
    mate-in-585.4%88.9%+3.5%
    mate-in-865.0%66.0%+1.0%

    Default play (UCI_AnalyseMode=false) is bit-for-bit unchanged —
    bench 13 @ Threads=1 produces exactly 1,107,765 nodes on both
    old and new binary.

  • NNUE EvalUseSmallOnly experiment (635e2ce) — tombstoned at
    -47 ELO; infra retained for future re-experimentation.

  • A4/A6 LTC SPSA retry (d192f25) — tombstoned at 0.0 ± 53 ELO;
    the 8 search/TM tunables are at a local optimum at all TCs.

Downloads

Pick the build that matches your CPU:

  • Hypersion-x86-64-avx2.exe — broadly compatible (Intel Haswell+ / AMD Zen+)
  • Hypersion-x86-64-bmi2.exe — uses PEXT for sliders (Intel Ice Lake+ / AMD Zen 3+)
  • Hypersion-x86-64-avxvnni.exe — uses 256-bit VNNI for NNUE inference (Alder Lake+ / Zen 4+), +29.6 ± 35.6 ELO over avx2 in this codebase

Three NNUE network files are required at the same directory as the
binary: nn-c288c895ea92.nnue (big), nn-37f18f62d772.nnue (small),
Perfect2023.bin (opening book). They're listed in the repo root.
The convenience .zip bundles include everything pre-packaged.

Status

Hypersion has reached a strong local optimum on its current SF18 NNUE
network. The most realistic remaining ELO ceiling is an NNUE
retrain (architecture + training data both available; project
author lacks GPU resources). Active development resumes when a
contributor steps up — open an issue if interested.



Comments

Post a Comment