8000
Skip to content

PokerBotAI/awesome-poker-ai

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

2 Commits
 
 

Repository files navigation

Awesome Poker AI Awesome

A curated list of AI poker research, open-source tools, solvers, and educational resources. Updated for 2026.

Poker is one of the most challenging domains for AI due to imperfect information, hidden cards, bluffing, and multi-player dynamics. This list covers the key research breakthroughs, open-source projects, and tools that have shaped the field.

Contents

Research Papers

Landmark Papers

  • DeepStack — First AI to defeat professional poker players in heads-up no-limit Hold'em using recursive reasoning and deep learning (University of Alberta, 2017). Published in Science.

  • Libratus — Defeated four top human pros in a 120,000-hand HUNL competition using blueprint strategy, real-time sub-game solving, and self-improvement (CMU, 2017).

  • Pluribus — First superhuman AI for six-player no-limit Hold'em, trained in 8 days on a 64-core server using Linear CFR (Meta + CMU, 2019). Author PDF.

  • ReBeL — General self-play RL + search framework that converges to Nash equilibrium in two-player zero-sum games; achieves superhuman HUNL poker (Meta, 2020). Official code.

Post-2020 Papers

  • Student of Games — Unified algorithm for both perfect and imperfect information games; achieves strong performance in chess, Go, and HUNL poker (DeepMind + Alberta, 2023). Published in Science Advances.

  • AlphaHoldem — End-to-end RL framework using a pseudo-siamese architecture; beats Slumbot and DeepStack after 3 days training on a single PC (Tsinghua, AAAI 2022).

  • DecisionHoldem — Open-source HUNL AI combining blueprint strategy and real-time safe depth-limited solving. GitHub.

  • PokerGPT — End-to-end lightweight solver for multi-player Texas Hold'em via fine-tuning a pre-trained LLM (OPT-1.3B) with RLHF on real game logs. Demonstrates that small language models can learn poker decision-making without expensive CFR computation (Huang et al., 2024).

  • SpinGPT — First LLM specifically trained for Spin & Go poker tournaments. Two-stage training: supervised fine-tuning on 320,000 expert decisions + reinforcement learning on 270,000 solver-generated hands. Achieves 78% solver agreement and 13.4 BB/100 vs Slumbot in heads-up play. Accepted at ACG 2025, published in LNCS (Maugin & Cazenave, 2025).

LLMs vs Poker: The 2025-2026 Wave

The explosion of general-purpose LLMs created massive public interest in whether ChatGPT, Claude, and Grok could play poker. Short answer: they can't — not at a competitive level. But the experiments were fascinating.

Events

  • Kaggle AI Poker Showdown (February 2026) — Google DeepMind and Kaggle organized the first major LLM-vs-LLM poker exhibition. Ten flagship models (GPT-5.2, GPT-o3, GPT-5 mini, Grok-4, Grok 4.1 Fast Reasoning, Gemini 3 Pro, Gemini 3 Flash, Claude Opus 4.5, Claude Sonnet 4.5, DeepSeek-V3.2) played 900,000+ hands of heads-up NLHE. GPT-5.2 was the most profitable overall; o3 won the exhibition bracket. All models played without solvers or CFR — pure text-pattern decisions. Doug Polk, Nick Schulman, and Liv Boeree provided commentary.

  • PokerBattle.ai (October-November 2025) — Five-day LLM cash game marathon organized by Max Pavlov. Nine models played Texas Hold'em around the clock. OpenAI o3 won with $36,691 profit over 3,799 hands. Meta LLAMA 4 performed worst, busting its entire $100K bankroll. Demonstrated the same fundamental LLM weaknesses: excessive aggression, poor bluffing comprehension, inability to fold.

Key Takeaways

The consensus from Kaggle, PokerBattle.ai, and professional analysis is clear:

  1. LLMs do not understand poker mathematics. They generate plausible-sounding text about poker but cannot reliably calculate equity, pot odds, or EV.
  2. LLMs hallucinate hand strength. They confuse suits, misread boards, and misidentify winning hands — even when they answer correctly in isolation.
  3. LLMs cannot adapt to opponents. No cross-hand memory, no range construction, no exploitative adjustments.
  4. Specialized poker AI (CFR-based) solved this years ago. Libratus (2017) and Pluribus (2019) achieved superhuman play through game theory, not language modeling. LLMs represent a step backward for poker AI, not forward.

SpinGPT is the notable exception — but it requires solver-generated training data and works only in a severely constrained format (3-player Spin & Go with short stacks).

Open-Source Frameworks

  • OpenSpiel — Google DeepMind's collection of environments and algorithms for RL research in games, with extensive poker support (Kuhn, Leduc, ACPC universal poker interface) and CFR/MCCFR implementations.

  • RLCard — Toolkit for RL in card games; supports Limit/No-Limit Hold'em, Leduc, Blackjack, Mahjong, UNO, and more. Website.

  • PokerRL — Multi-agent deep RL framework implementing NFSP, Deep CFR, Single Deep CFR, and RPG; supports distributed computing via Ray.

  • PyPokerEngine — Lightweight Python poker engine for AI development with an Emulator class for RL and a browser-based GUI.

  • PokerKit — Comprehensive Python library supporting Texas Hold'em, Omaha, Stud, Razz, and custom game variants (University of Toronto). Paper.

  • clubs — Python poker engine with OpenAI Gym interface; supports arbitrary community card game configurations (~714K hand evaluations/sec).

  • neuron_poker — Texas Hold'em OpenAI Gym environment with Keras-RL and a C++ equity module (~500x faster than Python).

  • deepcfr-texas-no-limit-holdem-6-players — Deep CFR implementation for 6-player NLHE with progressive training phases (random opponents → self-play → mixed pools), GRU-based opponent modeling, and a PyQt5 GUI. 80+ stars (2024-2025).

Hand Evaluators

  • PokerHandEvaluator — High-performance C/C++ evaluator using a perfect hash algorithm; supports 5-7 card hands and Omaha (PLO4/5/6). Python bindings available.

  • treys — Python 3 hand evaluator using bit arithmetic and lookup tables; evaluates 5/6/7 card hands (~250K evaluations/sec).

  • deuces — Original pure-Python hand evaluation library, written for the MIT Pokerbots Competition.

Open-Source Solvers

  • TexasSolver — Free open-source Texas Hold'em GTO solver with a GUI (Windows/macOS/Linux); performance comparable to PioSOLVER. Website.

  • postflop-solver — Rust-based postflop GTO solver using Discounted CFR. Also available as a web app and desktop app.

  • slumbot2019 — C++ CFR implementations (CFR+, MCCFR, Targeted CFR) by the creator of Slumbot, a multi-year ACPC champion.

  • ReBeL — Meta's official open-source implementation of the ReBeL algorithm for imperfect-information games.

Commercial Solvers

Name Website Description
PioSOLVER piosolver.com Industry-standard NLHE postflop solver
GTO+ gtoplus.com Fast, memory-efficient NLHE solver with multi-way support
MonkerSolver monkerware.com Industry standard for PLO and multi-way spots
Deepsolver deepsolver.com Cloud-based solver, solves any spot on-demand
HRC holdemresources.net Tournament (ICM) preflop analysis

Training Platforms

Name Website Description
GTO Wizard gtowizard.com AI-powered solver + trainer with on-demand solving. In 2025, partnered with GGPoker, WPN, and WPT Global for anti-RTA detection — tracking real-time solver use during play
PokerSnowie pokersnowie.com Neural network poker coach with play-against-AI mode

AI Poker Bots

Commercial

Historical

  • Slumbot — Multi-year Annual Computer Poker Competition champion by Eric Jackson; plays HUNL against humans for free online.

LLM Poker Projects

A growing number of projects attempt to use general-purpose LLMs for poker. None are competitive with CFR-based solvers, but they serve as research tools and benchmarks for LLM reasoning capabilities.

  • poker_LLM — AI-powered Texas Hold'em framework where LLMs (OpenAI, Claude, DeepSeek, QWen) act as players. Includes a Vue 3 replay visualization system and AI self-reflection on decisions. Python + Vue 3 (2024-2025).

  • HarperJonesGPT/PokerGPT — PokerStars screen-reading bot using GPT-4 API for decision-making via OCR + prompt engineering. Not fine-tuned — pure API calls. A hobby project, not competitive.

Benchmarks & Evaluation

  • poker-eval — TypeScript framework for evaluating AI agent performance in simulated NLHE cash games. Uses standard poker KPIs (BB/100, EV, VPIP). Supports Vercel AI SDK, OpenAI, Mastra, LlamaIndex, and Langchain. Includes a leaderboard of LLM performance (2024).

    Current leaderboard (1000 hands vs 2x GPT-4o baseline):

    Model BB/100
    mistral-large-latest +11.26
    gpt-4o -14.78
    claude-3-5-sonnet -19.95
    gpt-4o-mini -45.09
    gemini-1.5-pro -166.85

Educational Resources

Poker Math & Strategy (by PokerBotAI)

Getting Started

Notable Commentary & Analysis

Coverage and analysis of the LLM-poker intersection from in 6AD0 dustry voices:


Contributing

Contributions welcome! Please submit a pull request or open an issue to suggest additions.

License

CC0

About

A curated list of AI poker resources: research papers, open-source tools, bots, solvers, and strategy guides

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors

0