Six-Player No-Limit Hold'em AI
A 6-max poker AI plays no-limit Texas hold'em with six players at the table. It is a harder problem than heads-up, where one opponent faces one bot, and for years it was the open one. This article explains why multiplayer poker AI is difficult, how Pluribus approached it, what is available in the open today, and how to evaluate a six player poker AI fairly.
Why six players is harder than two
Heads-up poker is a two-player zero-sum game: what one player wins, the other loses. For these games there is a clean theory. A Nash equilibrium strategy cannot lose in expectation, and counterfactual regret minimization (CFR) provably converges toward one. Libratus and DeepStack built on this in heads-up no-limit.
With three or more players, the guarantees go away. Several things change at once:
- No equilibrium guarantee. A Nash equilibrium still exists, but playing your part of one does not protect you. If other players choose strategies from different equilibria, the combination can be bad for you. CFR is no longer known to converge to an equilibrium in this setting.
- Implicit collusion. Two opponents can, without communicating, play in a way that happens to hurt the third. A two-player proof cannot rule this out. Evaluation has to assume nobody colludes, and bots must not try to.
- More ranges to track. Each opponent has about 1,326 possible hole-card pairs, and every action updates your belief about all of them. The game tree also grows quickly: more players means more actions per street and more ways to reach a showdown.
- Position matters in six ways. A strategy has to be right for the button, the blinds and every seat between, and for any number of players still in the hand.
How Pluribus approached it
Pluribus, by Noam Brown and Tuomas Sandholm (Science, 2019), was the first AI to beat elite professionals at six-player no-limit hold'em. It did not try to solve the game. It made practical choices and let the evidence decide whether they worked.
- Blueprint. A strategy for the whole game computed offline with Monte Carlo CFR through self-play. The paper reports about 8 days on a 64-core server, roughly 12,400 CPU core hours. Cards and bet sizes are grouped by abstraction to keep the game small enough.
- Real-time search. During play, Pluribus refines the blueprint for the current situation with depth-limited search. At the leaves, each player picks among four continuation strategies: the blueprint and versions biased toward folding, calling and raising. This makes the search robust to opponents who change what they do.
- Hardware. Two CPUs and under 128 GB of RAM at play time.
It won in two formats: five professionals against one Pluribus over 10,000 hands, and one professional against five copies of Pluribus over 5,000 hands each for Darren Elias and Chris Ferguson. In the five-humans format the margin was about 48 mbb/game, roughly 5 big blinds per 100 hands, with statistical significance using AIVAT. The authors' code and trained strategy were never released. The Pluribus article has more.
What is open today: SHARINGAN-2
Poker Sharingan is an open, from-scratch reproduction of Pluribus in Rust (Apache-2.0), with an arena that scores six-player bots. SHARINGAN-2 is its model: a blueprint trained to a budget of 1.83e11 MCCFR iterations (finished November 2, 2026), played with real-time search on the turn and river. Weights are on Hugging Face and the code is on GitHub.
Here are the results that exist, stated as measured:
- An earlier checkpoint (2e10 iterations, with search) won +77.6 bb/100 over 2,329 live hands on a public bot arena.
- A later checkpoint (4.77e10 iterations) beat SvanBot, the top bot on that arena, in both seatings.
- The final SHARINGAN-2 beats the 2e10 checkpoint by +5.09 ± 1.16 mbb/game and the 4.77e10 checkpoint by +1.97 ± 0.97 mbb/game, in a 3 vs 3 seating over 4 million deals with duplicate and AIVAT (95% confidence intervals).
- A short best-response run (12M iterations) found no profitable exploit. That is a sanity check, not a measure of exploitability.
SHARINGAN-2 has not played professional humans, and no claim is made here that it is unexploitable or superhuman. It differs from the paper in two known ways: the exact Pluribus bet menus are unpublished, so three menus were built and the medium one won by measurement, and search runs on the turn and river only. You can play it in the browser.
How to evaluate 6-max bots
Six-handed results are noisy, and the field matters. A bot that beats five weak opponents may lose to five strong ones. Use more than one seating, and always correct for luck.
| Seating | Setup | What it tells you |
|---|---|---|
| 1 vs 5 | One test bot against five copies of a reference bot | Can the bot beat a fixed opponent? This is the leaderboard format against SHARINGAN-2. |
| 3 vs 3 | Three copies of bot A against three copies of bot B, seats alternating | Head-to-head strength between two bots. Used for the SHARINGAN-2 checkpoint comparisons above. |
| 5 vs 1 | Five copies of the test bot against one reference bot | Is the bot robust when it faces itself? Pluribus was also tested as one professional against five copies of itself. |
| Mixed field | Several different bots in one table | Closest to a real lobby, but results depend on who is in it. |
Then add the corrections. Duplicate replays the same deals with seats rotated, so every bot holds every set of cards. AIVAT ( Burch et al., 2018) uses a value function to subtract chance luck without bias. Report mbb/game with a 95% confidence interval. Measuring poker bots explains each step.
To test your own bot, the arena does this for you. Write act(s), run sharingan submit mybot.py, and read the verified score on the leaderboard. The poker bot guide walks through it, and the docs list the commands.
FAQ
Why is multiplayer poker harder for AI than heads-up?
Heads-up poker is two-player zero-sum, where equilibrium strategies have guarantees and CFR converges. With three or more players those guarantees no longer hold, there are more opponent ranges to track, and the game tree is much larger.
Has a 6-max poker AI beaten professionals?
Yes. Pluribus (Brown and Sandholm, Science 2019) beat elite professionals in two six-player formats. Its code and strategy were not released.
Is there an open source 6-max poker AI?
SHARINGAN-2 is an open reproduction of Pluribus with weights on Hugging Face and code on GitHub under Apache-2.0, and published results against its own earlier checkpoints. It has not been tested against professional humans.
What is a 6-max NLHE bot?
It is a program that plays no-limit Texas hold'em at a table of up to six players, choosing fold, check, call or raise each turn. You can write one in Python and score it against reference bots in the Poker Sharingan arena.
How do you measure a six-player poker bot?
Use duplicate deals so every bot sees every card set, apply AIVAT to remove more luck, run several seatings such as 1 vs 5 and 3 vs 3, and report a 95% confidence interval in mbb/game.