Poker AI History: Libratus, Pluribus, DeepStack, ReBeL
Poker AI history runs from limit hold'em solvers to superhuman six-player bots in about a decade. This page lays out the poker AI timeline, then compares Libratus vs Pluribus, DeepStack, ReBeL, Cepheus and Student of Games on game, players, method and what was released. It ends with the open reproduction, SHARINGAN-2.
Why poker was hard
Chess and Go are perfect-information games: both players see everything. Poker is not. Your opponent's cards are hidden, so the best action depends on what you think they hold, and what they think you hold. Good play also requires bluffing at the right frequency, which means the best strategy is randomized. That is a game-theory problem, not a pure search problem. Most of the systems below build on counterfactual regret minimization.
A third difficulty is that mistakes are hard to see. A bluff that fails and a bluff that was a bad idea look the same in one hand. Judging a poker AI needs thousands of hands and statistics, which is the subject of measuring poker bot skill.
Size is the second problem. Heads-up limit hold'em has about 3 x 10^14 information sets. No-limit has far more, because bets can be any size. All systems here shrink the game or approximate parts of it.
Poker AI timeline
| Year | System | Milestone |
|---|---|---|
| 2007 | CFR | Zinkevich et al. introduce counterfactual regret minimization (NeurIPS). |
| 2015 | Cepheus | Bowling, Burch, Johanson and Tammelin essentially weakly solve heads-up limit hold'em with CFR+ (Science). |
| 2015 | Claudico | Carnegie Mellon's heads-up no-limit bot loses to pros in the Brains vs AI match. |
| 2017 | DeepStack | Moravcik et al. (Science) use continual re-solving and neural value functions in heads-up no-limit. |
| 2017 | Libratus | Brown and Sandholm's bot beats four top pros over 120,000 hands in January (Science paper in 2018). |
| 2019 | Pluribus | Brown and Sandholm beat elite pros at six-player no-limit hold'em (Science). |
| 2020 | ReBeL | Brown, Bakhtin, Lerer and Gong combine reinforcement learning and search for imperfect-information games (NeurIPS). |
| 2023 | Student of Games | Schmid et al. at DeepMind present one algorithm for perfect and imperfect information games (Science Advances). |
| 2026 | SHARINGAN-2 | Open from-scratch Pluribus-style reproduction with released weights. |
Two threads run through the table. The first is the size of the game: limit, then heads-up no-limit, then six players. The second is the method: tabular CFR solutions, then CFR blueprints with real-time search, then learned value functions inside search. Later systems keep regret minimization as a component, and add search or learning around it.
The systems, briefly
Cepheus (University of Alberta, Science 2015) showed that a small game, heads-up limit, could be essentially solved. It used CFR+, which resets negative regret to zero. The result was a strategy a human cannot beat in a lifetime of play.
Libratus (Science 2018) moved to heads-up no-limit. It computed a blueprint with Monte Carlo CFR, then re-solved subgames in real time and patched weaknesses the opponents found. It beat four professionals over 120,000 hands in January 2017. Its predecessor Claudico had lost in 2015.
DeepStack took a different path. It never builds a full blueprint. At each decision it re-solves a depth-limited subgame, using neural networks to estimate the value of positions beyond the horizon. The idea of learned value functions inside search returns in ReBeL.
Pluribus is the multiplayer step. In two experiments, five professionals against one Pluribus, and one professional against five copies, it came out ahead. It did this with a CFR blueprint trained in 8 days on a 64-core server, plus depth-limited search at play time using four continuation strategies at leaf nodes. The code and trained strategy were never released. Our Pluribus overview goes deeper.
ReBeL generalizes AlphaZero-style self-play to imperfect information by searching over public belief states. Student of Games pushes further: one algorithm tested on chess, Go, poker and Scotland Yard. Both are heads-up in poker.
Libratus vs Pluribus and the others: a comparison
| System | Game | Players | Core method | Year | Code or weights released? |
|---|---|---|---|---|---|
| Cepheus | Limit hold'em | 2 | CFR+ tabular solution | 2015 | Playable online; no code or weights |
| Libratus | No-limit hold'em | 2 | MCCFR blueprint, subgame solving, self-improvement | 2017 | No |
| DeepStack | No-limit hold'em | 2 | Continual re-solving, neural value networks | 2017 | Small-game code only |
| Pluribus | No-limit hold'em | 6 | MCCFR blueprint, depth-limited search | 2019 | No |
| ReBeL | Poker, Liar's Dice | 2 | RL plus search on public belief states | 2020 | Liar's Dice code, no poker model |
| Student of Games | Several games incl. poker | 2 (poker) | Unified search with learning | 2023 | No poker weights that we know of |
| SHARINGAN-2 | No-limit hold'em | 6 | MCCFR blueprint, turn and river search | 2026 | Yes, Apache-2.0 and Hugging Face |
Read the "Players" column carefully. Libratus vs Pluribus is not a like-for-like comparison, since they play different games. The Libratus line of work led directly to Pluribus: same authors, same blueprint-plus-search shape, extended from two to six players.
Heads-up versus multiplayer
Two-player zero-sum games have a clean theory. A Nash equilibrium strategy cannot lose in expectation, and CFR converges to one. With three or more players, that guarantee disappears: playing an equilibrium strategy no longer guarantees you break even, and CFR is not guaranteed to find an equilibrium. Pluribus showed that self-play still works well in practice, which is why it was a milestone. See six-max poker AI for what that means for design.
Open reproductions
Almost every system above shipped as a paper, not as software. Open heads-up bots exist: Slumbot is a well-known public heads-up bot, and DecisionHoldem published a heads-up no-limit agent in 2022. Several GitHub repositories are named after Pluribus, but as far as we know none ships full-game trained six-player weights with published head-to-head results.
SHARINGAN-2 aims to fill that gap. It is a Rust reproduction of the Pluribus recipe, with weights on Hugging Face and code on GitHub. Its blueprint finished training at 1.83e11 MCCFR iterations and is played with real-time turn and river search. It has not played professional humans, so we make no claim about superhuman play. Scores come from a duplicate and AIVAT arena; see how poker bots are measured, the leaderboard, or play against it. The open-source Pluribus page lists what is and is not reproduced.
FAQ
What was the first poker AI to beat professionals?
It depends on the game. Heads-up limit hold'em was essentially solved by Cepheus in 2015. In heads-up no-limit, Libratus beat four top professionals in January 2017. In six-player no-limit, Pluribus beat elite professionals in 2019.
What is the difference between Libratus and Pluribus?
Libratus plays heads-up (two-player) no-limit hold'em, Pluribus plays six-player. Both use a Monte Carlo CFR blueprint plus real-time search, but Pluribus uses depth-limited search with four continuation strategies and runs on far less hardware.
How does DeepStack differ from Libratus?
DeepStack re-solves the game from the current state at every decision, using neural networks to estimate values past a depth limit. Libratus computes a large blueprint offline and refines it with subgame solving.
What is ReBeL?
A 2020 Facebook AI algorithm that combines reinforcement learning with search over public belief states for imperfect-information games, evaluated on heads-up poker and Liar's Dice.
What is Student of Games?
A 2023 DeepMind algorithm that plays both perfect-information games such as chess and Go and imperfect-information games such as poker with one method.
Is there an open-source Pluribus?
Pluribus itself was never released. SHARINGAN-2 is an open reproduction of the recipe with code and weights, though it has not played professional humans.