Onyx a UCI chess engine written from scratch in C++, with a self-trained NNUE evaluation. Estimated strength: ~3200.
Onyx was built and trained on a single desktop PC (Intel i7-11700K + RTX 3070) — the neural network's training data was generated entirely by the engine playing itself, with no external datasets or pretrained weights.
Author: Dylan2554 Rating Chess Engines Diary CEDR=3311
Onyx 2.0 what's new?
~3390 CCRL-Blitz-equivalent, measured: +62.9 +/- 17.1 Elo vs Stash 34.0 (CCRL Blitz 3328) over 1,000 games at 10s+0.1s, single thread.
The last public release was Onyx 1.6; versions 1.7-1.8 were internal development builds. Relative to the internal 1.8 baseline this release gains roughly +280 Elo at fixed nodes: a new NNUE (gen8, trained on 97.5M self-play positions via a free-cloud pipeline) plus 12 SPRT-gated search/eval changes. Every accepted and rejected experiment is documented in EXPERIMENT_LOG.md.
Downloads: grab the zip (recommended - both Windows builds, net, book, source, checksums), or the standalone files below. The exe requires onyx.nnue in the same folder - if you take a bare exe, take the net too. Use the avx2 build on any 2013+ CPU; the generic build runs anywhere.
Written entirely by AI agents (Anthropic Claude + OpenAI Codex) under human direction; all code original to this project. MIT.
Comments
Post a Comment