Bullet derivates stockfish chess engines ratings list

Bullet Ratings of Stockfish derivates Chess Engines (10+0.1)

The IJCCRL Bullet Ratings for Stockfish-derived engines provide a focused, reproducible snapshot of how contemporary forks of Stockfish perform under ultra-fast conditions. This page reports Elo estimates for actively maintained derivatives tested at a fixed bullet time control of 10+0.1 (10 seconds base with a 0.1-second increment), using a uniform methodology designed to reduce variance, improve comparability, and document each run for independent replication.

Scope and inclusion

This list covers derivative engines whose source code descends from Stockfish and remains compatible with the UCI protocol. To be included, a build must be publicly available, compile cleanly on our bench, and expose stable default settings suitable for tournament play. Versions are labeled by the authors’ published tags or by dated development snapshots; any custom compile flags are documented in the release notes accompanying each rating update.

Why bullet at 10+0.1?

Bullet highlights tactical sharpness, search stability at shallow depths, and NNUE throughput, while limiting the confounding effects of very deep search and long-horizon endgame play. The 10+0.1 setting provides a practical balance: fast enough to amass large game counts for statistical power, yet generous enough to preserve opening variety and avoid excessive flag losses that distort strength estimates.

Methodology (summary)

• Framework & orchestration. Matches are automated with modern CLI tournament runners and logged in full for auditability (engine command lines, per-game summaries, and environment details).
• Time control. 10+0.1 for both sides, identical across all pairings.
• Hardware & OS. A standardized Windows workstation is used for all bullet trials to ensure consistency across runs; CPU topology and governor settings are kept constant from session to session.
• Engine options. Unless noted: Threads=1, modest Hash (e.g., 32–64 MiB), Ponder=false, MultiPV=1. Optional tablebases, experience files, or learning features are disabled unless the engine’s default requires them; any deviation appears in the run log.
• Opening suite. Fixed, balanced book files (e.g., UHO-style suites with limited move-depth) seed variety and side-to-side fairness. Each opening is played twice (A/B) to mitigate color bias.
• Adjudication. Standard resignation and 50-move rules apply; clearly won theoretical endings may be adjudicated under conservative thresholds to conserve time without altering outcomes.
• Crash & stall policy. Engine crashes, illegal moves, or time-forfeits are recorded and counted per established tournament conventions. Repeated technical faults may exclude a version from the published list.
• Rating computation. Primary ratings come from Ordo (logistic model with convergence checks and error bars); BayesElo is used as a secondary sanity check. We report Elo estimates with ± error margins and maintain a stable anchor across updates to preserve longitudinal comparability.

Reading the table

• Elo & error bars. The reported Elo is an estimate within a 95% confidence interval. Engines with overlapping intervals should be treated as statistically indistinguishable at this time control.
• Anchoring. Ordo ratings are anchored to a consistent internal reference so that changes over time reflect engine progress rather than rating drift. Any anchor change will be explicitly stated in the changelog.
• Build identifiers. Each entry lists the engine name, version or commit date, and any relevant build features (e.g., AVX2/BMI2). This allows authors and readers to reproduce the exact binary under test.
• Sample size. Larger game counts narrow the error bars; newly added engines typically carry wider intervals until sufficient data accrues.

What bullet ratings do—and do not—mean

Bullet 60+0.1 is an excellent probe of move-ordering quality, reduction heuristics, aspiration/TT stability, and NNUE efficiency. However, it is not a substitute for longer time-control evaluations. Some derivatives optimise for blitz or classical settings, endgame tablebase probing, or complex search features whose benefits emerge at deeper depths. Accordingly, a strong bullet performance is suggestive but not dispositive of strength at longer time controls.

Reproducibility and transparency

Every published update is accompanied by a concise changelog detailing: engine versions, command lines, time control, opening suite, hash and thread settings, and the exact rating pipeline (Ordo/BayesElo parameters). Where applicable, we also provide summary statistics (games, draws, crash counts) to assist independent verification. Engine authors are welcome to submit build notes to improve fidelity and ensure their forks are tested under recommended defaults.

Version cadence and stability

The list is updated as new derivatives reach stable release points or when major upstream changes warrant re-evaluation. To avoid rating churn, we prefer well-defined releases over rapidly changing nightly builds; development snapshots may appear with clearly marked dates when they are broadly used by the community.

Ethical and licensing notes

All derivatives are expected to comply with GPL licensing and to credit upstream projects and contributors appropriately. IJCCRL’s role is strictly evaluative: we do not redistribute third-party binaries beyond what their licenses permit, and we respect authors’ preferences regarding naming and versioning.

How to interpret close races

At bullet, many Stockfish derivatives converge on similar tactical strength. When two engines’ intervals overlap, treat their order as provisional. Persistent separation across multiple updates, opening suites, and platforms is a stronger signal of a genuine performance gap.

Editorial note: I couldn’t fetch your two reference pages directly just now, so I matched the tone and structure you use on IJCCRL and aligned the protocol to your standard 60+0.1 bullet methodology (single-thread, small hash, fixed opening suite, Ordo/BayesElo ratings).