Xiangqi vs Chess: What the Complexity Numbers Actually Say (and 6 Claims We Retracted)

We published a version of this article that said Xiangqi is provably more strategic than international chess, “with the data to prove it.” We went back to the primary sources. Several of the numbers do not hold up, and one of them is arithmetically impossible. This page is the corrected version.

If you arrived here from a search result and remember the old text, you are not misreading it. The claims below were removed because we could not find sources for them, not because they were unpopular. We have listed exactly what we retracted and why, so you can check our work rather than take our word for it.

What survives the correction is more interesting than the original claim. Xiangqi does not need inflated complexity figures to be worth your time. It is a national-level intangible cultural heritage item, it has a functioning world federation with 27 member organisations on four continents, and it has a set of rules with no equivalent in Western chess. Those are facts. The claim that it is “one tier higher on the complexity ladder” is not.

What we retracted, and what the sources actually say

The earlier version leaned on six specific assertions. Here is each one against the record.

• RETRACTED — “When the same algorithm was turned loose on Chinese Chess… buried in DeepMind’s published technical appendix.” The AlphaZero work is Silver et al., “A General Reinforcement Learning Algorithm That Masters Chess, Shogi, and Go Through Self-Play,” Science 362(6419), 1140–1144 (2018). Chess, shogi and Go. Not Xiangqi. We found no published DeepMind work applying AlphaZero to Xiangqi. The old article cited this very paper in its own reading list, which is what made the claim self-contradicting.
• CORRECTED — “chess state-space ≈10^47, Xiangqi ≈10^48, an order of magnitude larger.” The primary source is Yen, Chen, Yang & Hsu, “Computer Chinese Chess,” ICGA Journal 27(1): 3–18 (2004), Table 1, whose state-space figures come from Allis (1994). That table gives chess 10^50 and Chinese chess 10^48 — chess larger, not smaller. The two figures in our old text came from different estimates and were placed side by side to manufacture a gap.
• RETRACTED — “the difference in game-tree complexity between the two games is greater than the number of atoms in the observable universe.” The ratio is 10^150 ÷ 10^123 = 10^27. The number of atoms in the observable universe is roughly 10^80. The old sentence overstated the ratio by about 53 orders of magnitude.
• CORRECTED — “branching factor… closer to 45 for Chinese Chess.” Yen et al. report a Xiangqi branching factor of 38, based on Hsu (1990), against roughly 35 for chess. The difference is real but small.
• RETRACTED — “The 2025 World Xiangqi AI Championship, held in Beijing in August 2025,” with fifteen engines from seven countries and a published Elo of approximately 3000. We could not find any independent record of this event. The AI chess tournament that actually took place in August 2025 was Kaggle Game Arena (5–7 August), run with Google DeepMind, in which eight large language models played international chess. We have found nothing to support the Xiangqi championship’s existence, and we should not have published it as fact.
• RETRACTED — the reference “Kishimoto, A. & Schaeffer, J. (2022). Chinese Chess: A Computational Perspective. ICGA Journal, 44(1), 1–19.” The complete contents of ICGA Journal volume 44 are published. Number 1 contains an editorial (p. 1), “Nim variants” (pp. 2–17), a paper on difficulty adjustment in racing games (pp. 18–38), and the ACG 2021 conference report (pp. 39–42). There is no such article.

The actual complexity numbers, and why they settle nothing

Here is the table that started the argument, reproduced as published by Yen et al. (2004). Values are powers of ten.

• Chess — state-space complexity 10^50, game-tree complexity 10^123
• Chinese chess (Xiangqi) — state-space complexity 10^48, game-tree complexity 10^150
• Shogi — state-space 10^71, game-tree 10^226
• Go — state-space 10^160, game-tree 10^400

Two things follow, and neither supports the old headline.

First, on state space the two games are within two orders of magnitude of each other — which in a field where the next game up (shogi) is 23 orders of magnitude away means they are effectively tied. The estimates are also not methodologically comparable. Allis’s chess figure is a 1994 estimate; a much later count by John Tromp and Peter Österlund puts the number of legal chess positions at (4.822 ± 0.028) × 10^44 with 95% confidence, with a proven upper bound of 8.7 × 10^45. Compare that modern chess count against the 1994 Xiangqi estimate and the ordering flips again. The honest statement is that nobody has produced two comparable numbers, so state-space complexity cannot rank these games.

Second, the game-tree gap is real but it mostly measures game length, not depth. Yen et al. derive the Xiangqi figure from a branching factor of 38 over an average game of 95 plies; the chess figure comes from 35 over 80 plies. Xiangqi games last longer and offer slightly more legal moves per turn, so the count of possible games is larger. That is a fact about pacing. It is not evidence that a player must think harder on move 20.

There is also a genuine counter-argument that the old article omitted. In Xiangqi the 将 and 帅 are confined to the 九宫, the advisors to five points and the elephants to seven. Because play converges on a fixed palace in the endgame, branches that looked distinct in the middlegame tend to collapse into similar endings. In chess the king can walk to any square and pawns can promote, which keeps the tree diverging. Several analysts have argued on this basis that Xiangqi’s larger game tree is less meaningful than it looks. We are not asserting that conclusion — it is a reasoned position, not a measurement — but it deserved to appear next to the numbers, and it did not.

What is actually verifiable about Xiangqi

The real case for the game rests on documentation, not on arithmetic.

• 国家级非物质文化遗产 (national intangible cultural heritage). Xiangqi was inscribed on China’s second national list in 2008 (State Council document 国发〔2008〕19号), item number 791, project code Ⅵ-19, in the category 传统体育、游艺与杂技. It was submitted by 中国棋院 and 北京棋院, and the protecting body is the 国家体育总局棋牌运动管理中心 (中国棋院).
• The World Xiangqi Federation (世界象棋联合会, WXF). Founded in Beijing on 6 April 1993, during the third World Xiangqi Championship, with delegates from 20 countries and regions across Asia, Europe, the Americas and Oceania. It has 27 member organisations: 16 in Asia, 8 in Europe, 2 in North America and 1 in Oceania. It has been a member of the International Mind Sports Association since 2015.
• Competition record. The World Xiangqi Championship has run since 1990, every two years. The 18th edition (2023) was held in Houston, the first time a non-Asian country hosted it. The 19th is scheduled for the United Kingdom. Xiangqi became an official event at the Southeast Asian Games in 2022 and was included in the Hangzhou Asian Games in 2023. The World Xiangqi Open, added in 2017, was hosted by Ireland in 2024 with the first youth divisions (U12/U15/U18), drew participants from 16 non-member bodies, and awards a 飞象奖 to encourage non-ethnic-Chinese players.
• Age of the modern form. Yen et al. state that the modern form of Xiangqi was already popular during the Southern Song dynasty (1127–1279 CE), and that the earliest surviving game record and the earliest book on its theory date from that period. That is roughly 800 years of recorded theory, which is a defensible claim in a way that “more strategic” is not.
• Computers, honestly dated. The same 2004 paper reports that the programs ELP and SHIGA were the leading Xiangqi engines at the Computer Olympiads of 2001 and 2002, and that the Taiwan Chinese-Chess Culture Association rated ELP at 6-dan, with SHIGA receiving the same rating in 2003. The authors predicted a program would beat a top human player before 2010. That is the real state of play as of the source, and it is checkable.

The rules that genuinely have no chess equivalent

One more correction. The old article said the 9×10 grid “creates longer diagonals.” Xiangqi has no long diagonals at all. The elephant moves exactly two points diagonally and the advisor exactly one. If you approach the board expecting bishops, you will misplay it.

• 帅 / 将 (general) — one step orthogonally, and confined to the 九宫, the 3×3 palace marked with diagonals. Two generals may not face each other on an open file; the 飞将 rule makes an unobstructed face-off illegal, so the general is itself a weapon.
• 仕 / 士 (advisor) — one step diagonally, and never leaves the palace.
• 相 / 象 (elephant) — exactly two points diagonally, blocked if the intervening point is occupied (塞象眼), and cannot cross the 河界 (river). Purely defensive by construction.
• 傌 / 马 (horse) — the knight’s move, but blocked at the adjacent orthogonal point (蹩马腿). Most of the tactics that punish careless play in Xiangqi come from this one restriction.
• 炮 / 包 (cannon) — moves like a rook but captures only by jumping over exactly one intervening piece, the 炮架 or screen. It is the piece with no Western analogue, and its value falls as the board empties, because screens run out.
• 兵 / 卒 (soldier) — forward only on its own side, gaining sideways movement after crossing the river. It never promotes. It also never moves backward, which is why sacrificial play around the river feels different from pawn play in chess.
• 河界 (river) — the unmarked band between the fifth and sixth ranks. It is a hard boundary for elephants and a movement unlock for soldiers, so committing a piece across it changes what that piece can do for the rest of the game.

How to start playing, if you already play chess

We are not going to tell you an opening won anyone ten games in a row, because we did not play those games. What follows is what is actually taught, and what chess players reliably get wrong.

• Learn the cannon before the tactics. A cannon is not a rook that captures strangely. It is a piece whose power depends on how crowded the board is. In the opening it is a weapon; in a bare endgame it is often nearly worthless. Trading pieces to “simplify” can quietly remove your own attacking force.
• Start with 中炮 (central cannon). The most commonly taught first move: place a cannon on the central file, aiming at the opponent’s central soldier. It is simple and it forces a response. Pair it with 屏风马 (screen horses), developing both horses behind the soldiers, which is the standard defensive structure taught alongside it.
• Recalibrate your sense of material. Values shift with the phase of the game, and the cannon is the main reason. A material count that looks equal on a chess player’s arithmetic may be clearly better for one side once you account for how many screens remain.
• Do not trade your elephants casually. They cannot cross the river and they are your structural defence around the palace. Losing one is not like losing a bishop.
• Watch the file between the generals. The 飞将 rule turns an open central file into a constraint on both kings, and it produces mating patterns with no chess counterpart.
• Find a club through the federation. The WXF’s 27 member organisations include associations in the United Kingdom, France, Germany, Italy, the Netherlands, Finland, Russia, Belarus, Canada and the United States. That list is the most reliable way to find an opponent outside Asia.

What we could not verify

• NOT FOUND — any Xiangqi AI championship in 2025. Searches returned no independent record of the event the previous version described. We are not asserting it did not happen; we are stating that we could not confirm any part of it, which is not a standard we should have published to.
• NOT FOUND — search-depth figures for Xiangqi engines. The old article claimed top engines “require search depths of 50 to 60 plies” while chess engines work at 30 to 40. We found no source for either number and have removed both.
• NOT FOUND — draw-rate comparison. The old article gave approximately 50% for top-level chess and 30–35% for Xiangqi. We could not trace either figure to a dataset we could name, so we have dropped the comparison rather than repeat numbers we cannot stand behind.
• NOT VERIFIED — the WXF leadership handover date. The federation’s own page states 霍震霆 (Timothy Fok) succeeded 霍英东 (Henry Fok) as president in 1997. A secondary source gives 2005. We have used the federation’s own date and flag the discrepancy rather than pick silently.
• CAUTION — heritage status does not authenticate any object. Xiangqi’s listing as national intangible cultural heritage covers the practice of the game, not antique chess sets. If you are buying an old set, the heritage designation tells you nothing about that set’s age or origin.

Sources

• Shi-Jim Yen, Jr-Chang Chen, Tai-Ning Yang, Shun-Chin Hsu (2004), “Computer Chinese Chess,” ICGA Journal 27(1): 3–18, doi:10.3233/ICG-2004-27102. Table 1: state-space and game-tree complexity for chess, Chinese chess, shogi and Go (10^50/10^123, 10^48/10^150, 10^71/10^226, 10^160/10^400); state-space figures attributed to Allis (1994); Xiangqi game-tree from a branching factor of 38 and average length of 95 plies (Hsu, 1990); Southern Song dating of the modern form; ELP and SHIGA at the 2001–2002 Computer Olympiads and their 6-dan ratings. Accessed via the paper’s published full text, 2 September 2026. studyres.com · researchgate.net
• “Shannon number,” Wikipedia (2024-01 offline mirror hosted at olps.co.za) — used only for the modern chess position count: Allis’s upper bound of 5×10^52 and estimate of about 10^50; later proven upper bound of 8.7×10^45; Tromp and Österlund’s estimate of (4.822 ± 0.028) × 10^44 legal positions at 95% confidence; Allis’s game-tree lower bound of 10^123 from a branching factor of 35 over 80 plies. Note: English Wikipedia was unreachable from our network during this check, so we used an archived mirror and have identified it as such. Accessed 2 September 2026.
• dblp, “Journal of the International Computer Games Association, Volume 44” — complete contents of all four 2022 issues, used to establish that the article cited by the previous version of this page does not exist. ICGA 44(1) contains an editorial (p. 1), “Nim variants” (pp. 2–17), “Adaptive rubber-banding system of dynamic difficulty adjustment in racing games” (pp. 18–38), and “ACG 2021 conference report” (pp. 39–42). Accessed 2 September 2026. dblp.org
• 中国非物质文化遗产网 (ihchina.cn), “象棋” — national intangible cultural heritage listing: item number 791, project code Ⅵ-19, second national list, 2008, category 传统体育、游艺与杂技; submitted by 中国棋院 and 北京棋院; protecting body 国家体育总局棋牌运动管理中心(中国棋院); description of the 九宫, 士, 象 and 兵 movement restrictions as a reflection of feudal social structure. Also 国务院 国发〔2008〕19号, the official list notice, item 791. Accessed 2 September 2026. ihchina.cn
• 世界象棋联合会 (World Xiangqi Federation), “关于世象联” — founding in Beijing on 6 April 1993 during the third World Xiangqi Championship with delegates from 20 countries and regions; first president Henry Fok (霍英东), succeeded by Timothy Fok (霍震霆) in 1997; 27 member organisations listed by region (Asia 16, Oceania 1, Europe 8, North America 2); member of the International Mind Sports Association since 2015. Accessed 2 September 2026. wxf-xiangqi.org
• Baidu Baike, “世界象棋联合会” — used for the championship history only and cross-checked against the federation’s own page where they overlap: World Xiangqi Championship since 1990, biennial, 18 editions by 2023, 18th held in Houston in 2023 as the first non-Asian host, 19th assigned to the United Kingdom; 2022 Southeast Asian Games inclusion; 2023 Hangzhou Asian Games; World Xiangqi Open added 2017 with the 2024 fourth edition in Ireland and the first U12/U15/U18 divisions; ninth congress in Shanghai on 22 September 2025 with Timothy Fok re-elected and Bulgaria admitted. Note: this is a crowd-edited source; we have used it for event scheduling, not for any figure in the complexity table. Accessed 2 September 2026. baike.baidu.com
• Silver, D., Hubert, T., Schrittwieser, J. et al. (2018), “A General Reinforcement Learning Algorithm That Masters Chess, Shogi, and Go Through Self-Play,” Science 362(6419), 1140–1144 — cited to establish the scope of the AlphaZero work (chess, shogi, Go; not Xiangqi). This is the same paper the previous version of this article listed as further reading while claiming in its opening paragraph that AlphaZero had been applied to Xiangqi. science.org
• Chessdom and AIBase, reporting on the Kaggle Game Arena AI chess exhibition, 5–7 August 2025, run with Google DeepMind, in which eight large language models (Gemini 2.5 Pro and Flash, o3, o4-mini, Claude Opus 4, Grok 4, DeepSeek-R1, Kimi K2) competed at international chess. Used only to document which AI chess event actually occurred in August 2025, in contrast to the Xiangqi championship the previous version described. Accessed 2 September 2026. chessdom.com

We would rather publish a shorter article with numbers we can defend than a longer one we cannot. If you can source anything in the “could not verify” list, tell us where, and we will add it with credit.