-
-
Notifications
You must be signed in to change notification settings - Fork 195
Playing strength
Fairy-Stockfish outperforms most chess variant engines, except for variants where many highly specialized engines exist (e.g., Shogi), and is the strongest engine in a large variety of variants such as Janggi, Sittuyin, Capablanca chess, etc. For most of the popular variants NNUE networks are available that can further improve playing strength, see the list of NNUE networks. When using NNUE evaluation Fairy-Stockfish plays at a superhuman level in all major regional chess variants, such as Xiangqi, Shogi, Janggi, and Makruk. Even in user-defined variants the engine has never encountered before Fairy-Stockfish usually plays at master level without any adaptations.
For standard chess, functionality is almost identical with official Stockfish, but the slowdown (>2x) due to overhead for fairy pieces and variants leads to >100 Elo weaker performance. When using NNUE the speed difference is lower than with classical evaluation, since the variant code has much less impact on NNUE than on classical evaluation. Actually, NNUE evaluation even is faster than classical, which is why Fairy-Stockfish uses pure NNUE instead of hybrid evaluation.
For the variants supported by lichess and multi-variant Stockfish (Crazyhouse, antichess, 3check, etc.), playing strength without NNUE is on par (+-200 Elo) with multi-variant Stockfish, and above other variant engines. Strong NNUE networks are available for all lichess variants and surpass all other engines.
For chess variants with fairy pieces (e.g., makruk, sittuyin, etc.), playing strength is above all other known multi-variant engines (e.g., Sjaak II, FairyMax, Nebiyu) and at least on par with the best (non-Stockfish) engines in the respective variant, except for Shogi, and Xiangqi, and Shatranj, see below.
When using NNUE evaluation Fairy-Stockfish plays on superhuman level, at around 3800 in the shogi engine rating list. When using the handcrafted/classical evaluation, playing strength is more than 1000 Elo lower, estimated around 2700. Compared to human players this should be on top amateur to professional level. For the top specialized shogi engine based on Stockfish, see YaneuraOu.
In shogi variants (minishogi, euroshogi, etc.), it seems to be the strongest available engine, e.g., being 3x winner of the UEC cup for Minishogi. Furthermore, it is able to practically solve small-sized shogi variants such as dobutsu, micro, and kyoto shogi.
When using NNUE evaluation, Fairy-Stockfish plays Xiangqi at superhuman level and is only a few hundred Elo below spezialized top engines. Using classical evaluation playing strength is slightly above the level of Elephant Eye and Cyclone 0.55, at least on master level. For a Stockfish fork specialized to Xiangqi inspired by Fairy-Stockfish NNUE see Pikafish.
With NNUE evaluation Fairy-Stockfish in Janggi is the strongest available engine and also stronger than pro players. Without NNUE playing strength still is at least on par with professional players as well as strong engines (such as Janggidosa), see e.g. https://www.youtube.com/playlist?list=PLGE6rVolMRXZwprT9CQYz3CAvU2Y5zNnj.
When no statistical uncertainty is specified, the estimate is derived from very few games (<100).
For test results marked with * a (usually much) stronger NNUE network exists, see the variant NNUE list, but the given test result is for classical/handcrafted evaluation.
| variant | relative elo | reference engine |
|---|---|---|
| chess | -109.26 +-9.7 * | MV-SF |
| crazyhouse | -83.94 +-15.5 * | MV-SF |
| giveaway | +76.06 +-14.2 | MV-SF |
| atomic | -268.47 +-17.3 * | MV-SF |
| 3check | +278.24 +-20.3 * | MV-SF |
| king of the hill | -87.63 +-15.0 * | MV-SF |
| racing kings | -32.05 +-12.1 * | MV-SF |
| horde | -132.54 +-16.3 * | MV-SF |
| losers | +53.58 +-21.0 * | MV-SF |
| extinction | +600.90 +-44.5 * | MV-SF |
| placement | +748.80 +-68.8 * | MV-SF |
| seirawan | -109.83 +-19.4 * | Seirawan-Stockfish |
| shatranj | -0.69 +-12.4 * | Shatranj-Stockfish |
| makruk | +230.16 +-16.7 * | Makruk-Stockfish |
| makruk | +300 +-100 * | Bilis v1.0 |
| makruk | +400 +-100 * | NebiyuChess 1.45 |
| makruk | >+400 * | Sjaak II |
| shatranj | -82.1 +- 32.4 | Stormphranj |
| shatranj | >+400 * | Tiyaga v1.0 |
| shatranj | >+400 * | Sjaak II |
| shatranj | >+400 * | NebiyuChess 1.45 |
| ASEAN | >+400 * | Sjaak II |
| sittuyin | >+400 | Sjaak II |
| shatar | >+400 | Sjaak II |
| losalamos | +600 * | Sjaak II |
| losalamos | +700 * | NebiyuAlien 1.45 |
| minishogi | >+400 * | Sjaak II |
| minishogi | +400 * | Crazywa |
| minishogi | +250 * | Shokidoki ICGA15 |
| minishogi | +250 * | Lima v2-00 |
| minishogi | +200 * | Lima v4 |
| breakthrough | >+400 * | GameMaster 2.0 |
For variants that are also supported by the normal version (i.e., variants on boards <= 8x8), using the version for large boards decreases playing strength by 50-200 Elo due to lower speed and a few functional differences. Playing strength estimates for variants with large boards are given below.
| variant | relative elo | reference engine |
|---|---|---|
| xiangqi | +100 * | Cyclone 0.55 |
| xiangqi | +100 * | Elephant Eye 3.31 |
| xiangqi | +300 * | Sjaak II |
| xiangqi | >+500 * | MaxQi |
| shogi | -400 * | apery wcsc26 |
| shogi | 0 * | Gikou 2 D9 |
| shogi | +300 * | apery (no eval file) |
| shogi | +300 * | Shokidoki ICGA15 |
| shogi | >+400 * | Crazywa |
| shogi | >+400 * | Sjaak II |
| capablanca | +700 * | NebiyuChess 1.45 |
| capablanca | +700 * | Sjaak II |