Skip to content
Home » IJCCRL Elite Original Chess Engines bullet tournament

IJCCRL Elite Original Chess Engines bullet tournament

The current IJCCRL Blitz Chess Engines Rating List is now based on a renewed Elite Original Engines tournament configuration, aligned with our 2025 Cup Season IJCRL NNUE Bullet Elite framework. This cycle replaces the previous setup while preserving the same methodological rigour and emphasis on independently-developed, author-maintained engines. The event is conducted with cutechess-cli under the label “Cup season Ijcrl 2025”, and is hosted on a dedicated local test bench rather than a shared or virtualised environment.

The present elite tournament includes the following original engines and binaries:
Alexandria 8.1.12 AVX2, Berserk 13 AVX2-PEXT, Dragon 3.3 64-bit, Integral AVX2, Lc0, Obsidian 1.60 AVX2-PEXT, PlentyChess 7.0.0 FMA, Stockfish x86-64 AVX2, and Caissa 1.23 x64 AVX2. Each engine is run with the official or author-recommended AVX2 / PEXT / FMA3-optimised binary where available, ensuring that modern CPU features are used consistently while keeping the playing conditions comparable across the field. All engines use 1 search thread and a fixed 32 MB hash allocation, so that scalability differences in multi-threading do not distort the resulting ratings.

The tournament itself is a round-robin competition managed by cutechess-cli, with multiple full cycles to reduce statistical noise. Pairings are generated as a round-robin with colour-reversed games (-tournament round-robin -games 2), and the schedule is extended over 15 rounds with repeated cycles (-rounds 15) to accumulate a substantial number of games for rating purposes. The option -repeat 2 is used in conjunction with the opening file so that each starting position is played twice with colours swapped, enhancing fairness and reducing opening-bias. Intermediate Elo estimates are produced at regular intervals (-ratinginterval 10), though final ratings for publication are computed only once all games have been completed and thoroughly checked.

All games start from curated positions taken from the UHO_2024_8mvs_big_+100_+119 suite, supplied as a PGN file
..\Openings\PGN\UHO_2024_8mvs_big_+100_+119.pgn. These balanced, human-curated positions are served in random order (order=random) and are designed to avoid heavy opening-book theory, early repetition, and trivial drawing lines. By forcing engines into rich middlegame structures after eight moves, the tournament focuses on search and evaluation quality rather than book preparation. The option -maxmoves 200 ensures that excessively long endgames are adjudicated consistently, avoiding pathological marathons that could bias the statistics.

Endgame play is further standardised by enabling Syzygy tablebases (-tb C:\Syzygy -tbpieces 6). This configuration gives all engines access to up to 6-man Syzygy bases, ensuring perfect play in a wide class of endings and avoiding arbitrary adjudication. Tablebase access is identical for every participant, so any residual differences in endgame performance reflect the engines’ search strategies and evaluation heuristics rather than unequal access to external data.

A key innovation in the current cycle is the hardware platform. The entire tournament is run on a modern mobile high-performance system based on an
AMD Ryzen 7 8745HS processor, paired with 32 GB of DDR5 RAM and a 1 TB NVMe SSD. This configuration offers high single-thread performance, excellent memory bandwidth, and very low I/O latency. The use of an NVMe SSD essentially eliminates storage bottlenecks when accessing Syzygy tablebases and opening files, while DDR5 memory reduces the risk of bandwidth saturation in hash-intensive positions. By fixing the hardware in this way and avoiding background tasks, IJCCRL ensures that the resulting ratings are stable, reproducible and directly attributable to engine strength rather than to operating-system noise or resource contention.

All games from the event are saved to a single PGN archive
(..\Games\games.pgn) for independent verification and future research. Ratings derived from this tournament feed directly into the bullet Chess Engines Rating List, where they are processed using established Elo calculation tools and published alongside confidence intervals and auxiliary statistics. In this way, the updated configuration preserves the tradition of IJCCRL’s earlier elite tournaments while taking full advantage of contemporary hardware and NNUE-optimised binaries, offering readers and subscribers a transparent and methodologically sound benchmark of today’s strongest original engines.

IJCCRL Elite Original Engines — Bullet 2026 — Specification
Launcher-enforced parameters (10s+0.1s, scientific & reproducible)
TC 10+0.1 Threads 1 Hash 32 MB UHO 2024 Syzygy 6-man Concurrency 2 OwnBook false
Scope: Elite original engines, full 2026 bullet season (840 games/engine at 10s+0.1s)
Goal: Scientifically curated bullet rating list for serious opening preparation and engine testing
IJCCRL Elite Original Engines — Bullet 2026 — Participants
Original engines under identical ultrabullet settings (10s+0.1s, 1 thread, 32 MB hash)
#EngineCategoryNotes
1stockfish 17.1Reference baselineAnchor engine for the elite original list at 10s+0.1s
2Alexandria-8.1.12-avx2Original engineCompeting under identical launcher settings and hardware
3PlentyChess-7.0.0-windows-fmaOriginal engineBullet / ultrabullet stress-test at 10s+0.1s
4caissa-1.23-x64-avx2Original engineSame UHO 2024 opening suite, same Syzygy configuration
5Obsidian160-avx2-pextOriginal engineStrictly identical testbed, no private books (OwnBook=false)
6dragon-3.3-64bitOriginal engineEvaluated as part of the same elite bullet pack
7berserk-13-avx2-pextOriginal engineUses identical options and launcher policy for comparability
8integral_avx2Original engineIncluded as part of the elite original engines 2026 bullet set
Fairness: same UHO 2024 suite, same options, same hardware, no private books
Composition: elite original engines anchored by Stockfish 17.1 as a baseline reference
IJCCRL Elite Original Engines — Bullet 2026 — Tournament Settings
Technical rules enforced by the launcher (transparent and reproducible)
ParameterValuePurpose
Time control10s + 0.1sUltrabullet profile: strong time pressure, testing search efficiency and stability
Threads1Ensures strict comparability (no multi-thread scalability issues)
Hash32 MBUniform transposition table allocation for all engines
Opening suiteUHO 2024 (PGN)Controlled opening diversity, colour-balanced and reproducible
OwnBookfalseDisables private opening books; focuses on engine evaluation & search
Syzygy6-manImproves endgame correctness, adjudication and clean conversion
Concurrency2Balanced throughput: enough games without overloading the hardware
Games per pairing2Colour symmetry: each pairing is played with both colours
Max moves200Prevents runaway games and keeps results statistically clean
Season2026Defines the scope of this bullet dataset for IJCCRL publications
Method: same rules for all engines, launcher-enforced settings, documented and reproducible
Outcome: coherent bullet ratings that complement classical and blitz IJCCRL lists

IJCCRL Elite Original Engines — Bullet Tournament 2026
Final standings — Combined Ordo & BayesElo snapshot (1680 games each)
#EngineOrdo ratingOrdo pointsOrdo gamesOrdo score %BayesElo+BayesElo gamesBayesElo score %Opp. diffDraw %
1stockfish 17.13771.41029.5168061651212168061%-941%
2Alexandria-8.1.12-avx23745.6962.0168057421212168057%-640%
3PlentyChess-7.0.0-windows-fma3733.8930.5168055321212168055%-537%
4caissa-1.23-x64-avx23709.2864.516805191212168051%-137%
5Obsidian160-avx2-pext3698.2835.0168050-21212168050%038%
6dragon-3.3-64bit3668.7756.0168045-291212168045%440%
7berserk-13-avx2-pext3651.2709.5168042-441212168042%641%
8integral_avx23621.9633.0168038-711212168038%1039%
White advantage: 0.00
Draw rate (equal opponents): 50.00 %