Repository navigation
Releases
CeylonDemon 3.3
CeylonDemon 3.3 is the strongest CeylonDemon so far.
Downloads
CeylonDemon-3.3-x86-64-avx2.exeworks on most modern computers. Use this one if you're not sure.CeylonDemon-3.3-x86-64-bmi2.execan be a little faster on newer CPUs (Intel Haswell and later, AMD Zen 3 and later).CeylonDemon-3.3-windows-x86-64.zipcontains both, plus README, NOTICE, LICENSE and the release notes.SHA256SUMS-3.3.txtlets you check the downloads.
What's new since 2.0
- Resonance v12 dual-frame network, trained from scratch on 400,746,429 positions of CeylonDemon's own self-play.
- Many search improvements from 2.2 to 3.3. In 3.3: singular extensions from depth 6, and late-move reductions for checks and captures.
Load the .exe into any UCI chess GUI (Arena, Cute Chess, BanksiaGUI, ChessBase).
Full details: RELEASE-3.3.md and CHANGELOG.md.
CeylonDemon 2.0
CeylonDemon 2.0 is a free UCI chess engine with Resonance, a dual-frame NNUE
evaluator that reads every position through two feature transformers — one
anchored on the side to move's own king, one on the opposing king — and reports
the residual between them as an uncertainty signal the search consumes directly.
Powered by the NARC Engine.
Downloads
| File | Requires |
|---|---|
CeylonDemon-2.0-x86-64-avx2.exe |
x86-64 with AVX2 — portable release build, use this if unsure |
CeylonDemon-2.0-x86-64-bmi2.exe |
x86-64 with AVX2 + BMI2 — Skylake and later, fast PEXT |
The neural network is linked into both executables, so each is a single
self-contained file — there is nothing else to download and no network file to
configure.
Strength
Roughly 3210–3350 on the CCRL Blitz scale, centred near 3280.
That is a cross-list approximation, not a measurement — CeylonDemon has never
played a CCRL-conditions game and appears on no CCRL list. It is derived from
the measured result below by the fit CCRL_Blitz ≈ 0.765 × CCI_STC + 1199
(RMS residual ±33), widened for unquantified systematics. The fit rests on one
weak calibration point and is applied 42 points below the range it was
calibrated over. It refers to the Blitz list specifically; no equivalent is
offered for CCRL 40/15, for which no calibration was done.
The measured result
2719.5 ± 32.8 — a locally transferred CCI-STC estimate under 10+0.1
conditions. Note the scale: this is the CCI STC list, not CCRL, and the two
are not interchangeable.
This is not an official CCI rating and not universal Elo. It is an
interpolation against a fixed set of anchor engines.
| Games | 320, all completed — no stalls, crashes or forfeits |
| Record | 39–82–199 (80.0 / 320, 25.0%) |
| Conditions | 10+0.1 s, 1 thread, 64 MB hash, UHO openings, no adjudication |
| Opponents | 8 anchors from 2761 (Monolith 3.0) to 3101 (Rose 1.0.0) |
| Tools | FastChess 1.8.2-alpha, Ordo 1.2.6 (-F 95, 10,000 simulations) |
| Hardware | Intel Core i7-6600U @ 2.60 GHz, 2C/4T |
The engine sits inside the anchor range rather than below it — outscoring
Monolith 3.0 at 43.8% and losing to everything above — so the rating is an
interpolation rather than an extrapolation. Seven of the eight per-opponent
residuals fall between −0.6 and +3.5 points. The exception is Rose 1.0.0, against
which 2.0 scored 0.5/40 where its rating predicts about 4.0/40; that is a genuine
outlier outside sampling variation and the cause is not yet identified.
Caveats. The ±32.8 is statistical error only; the error introduced by
transferring against these particular anchors is systematic and unquantified.
The result says nothing about other time controls, thread counts or hash sizes,
and version gaps under roughly 65 Elo are unresolvable at this precision.
CeylonDemon has never played a game under CCRL conditions and does not appear on
any public rating list. Independent testing is welcome.
Which build was measured. The rating run used the AVX2 build. The BMI2 build
is compiled from the same source and differs only in instruction selection; it
has not itself been through a rating run. Both builds were verified to produce
identical evaluations and identical best moves on the positions checked below.
Verification
CeylonDemon-2.0-x86-64-avx2.exe
SHA-256 ab7fb2eab6012c380873a1b416ba499b3cbe7172635b539dc615b8160fa3d979
Size 13,184,680 bytes
CeylonDemon-2.0-x86-64-bmi2.exe
SHA-256 37912ebb5844ced69c3711bb3959801b53b08ecf09717ab3f4e8eb6e66d3033b
Size 13,183,656 bytes
Both binaries are compiled from the source at this tag with GCC 15.1.0
(MSYS2 UCRT64), statically linked. Move generation was verified against the
published perft references before release:
| Position | Depth | Nodes |
|---|---|---|
| Start position | 5 | 4,865,609 |
| "Kiwipete" | 4 | 4,085,603 |
The AVX2 binary differs from the specific executable used in the rating run in
4 bytes only — the PE build timestamp and the resulting header checksum — with
identical code, data and embedded network.
Engine details
- UCI options:
Hash(128),Threads(1),Move Overhead(30),EvalFile
(<embedded>) - Resonance v10 evaluator, integer inference throughout, embedded network
- Lazy SMP, 1–256 threads
Provenance
The embedded weights were warm-started from a Stockfish teacher network. The
engine source contains no Stockfish code, but the network carries Stockfish
ancestry and is not presented as independently derived. A clean-provenance
network is under development. See NOTICE.md for full attribution, including
the search techniques the engine implements but did not originate, and the
disclosure of AI-assisted development.
Licence
GNU General Public License v3 or later. Copyright © 2026 Madushan Dissanayake.