An open-source no-limit hold'em solver with a local browser interface. Build custom preflop scenarios, study heads-up postflop play, model opponents, and compare strategies across hundreds of flops.
Runs on your computer with a Rust engine and optional NVIDIA CUDA acceleration. The browser UI needs no frontend build step.
Install Rust and Git, then clone the repository:
git clone https://github.com/MatthewPDingle/GTOpen.git
cd GTOpenUse the Rust MSVC toolchain and install the Desktop development with C++ workload from Visual Studio Build Tools.
.\GTOpen.cmdThe launcher builds the app, starts a local server, and opens your browser. If GTOpen is already running, it reopens that server and preserves its sessions.
For NVIDIA GPU acceleration, install the CUDA runtime compiler once before launching (requires Python/pip and an NVIDIA driver):
python -m pip install --target .cuda-nvrtc nvidia-cuda-nvrtc-cu12
.\GTOpen.cmdInstall a C/C++ build toolchain, then run:
./start.shOpen http://127.0.0.1:3737. The launcher selects a CUDA build when it detects NVIDIA support; otherwise it builds for CPU. For CUDA without a full toolkit:
python3 -m pip install --target ~/.local/cuda-nvrtc nvidia-cuda-nvrtc-cu12
./start.shFor a manual CPU build, run cargo build --release -p server, then ./target/release/gto-server from the repository root.
The first build takes longer. Large trees need substantial RAM; GPU solving also depends on available VRAM. The app estimates memory before solving and can fall back to CPU.
| Tab | What to do |
|---|---|
| 00 · Preflop Lab | Build a 2–9-player game, choose blinds and raise sizes, assign optional player models, and solve. Follow an action line to a heads-up flop, then Send to postflop setup. |
| 01 · Setup | Enter or import both ranges, choose the board, and set pot, effective stack, rake, and bet sizes. |
| 02 · Solve | After Build tree in Setup, solve the selected board. Watch convergence, stop/resume, and save the session. |
| 03 · Browse | Follow the action ribbon and inspect strategy, EV, equity, and individual hands. Lock frequencies or a player model, then re-solve to study the response. |
| 04 · Reports | Solve the Setup configuration across 47, 95, 184, or all 1,755 canonical flops. Compare results by board, high card, hand class, or texture. |
You can start directly in Setup if you already have ranges. Preflop Lab is optional.
Units: preflop stacks and sizes are in big blinds; Small blind and Big blind specify the stake amounts (enter 2 and 2 for $2/$2). Postflop bet sizes are percentages of the pot; 33 75 offers two sizes, a means all-in, and 2.5x means a raise multiple.
Preflop sizes: expand Sizes by position to override opens, 3-bets, and 4-bets or later raises for each seat. Blank cells inherit the defaults above. A 3-bet multiplier of 4 facing a 6bb open raises to 24bb. Build a fresh game after changing sizes.
Preflop rake: new scenarios use Balanced (approx.), which splits one pot between the heads-up players and deducts your configured rake explicitly. Set rake to 0 for no rake; with positive rake, a cap of 0 means uncapped. Old saves retain their old model. How the approximation works.
For a normal live straddle, enable UTG straddle in Preflop Lab and enter its size (default 2 bb, tables of 3–9 players). The next seat acts first and the straddler keeps the final check/raise option. All sizes stay in the original big blind: a 2 bb straddle needs an ordinary opening raise of at least 4 bb. Save the setting with your scenario or game. Straddle rules and model limitations.
Use Manage models to create, edit, copy, or remove models from the menus. Assign models to seats, then apply them and re-solve. The Evidence indicator distinguishes measured hand policies and sizing from estimates and fallbacks.
The library includes Ignition NL10 and CoinPoker NL10/NL25/NL50/NL100 models. Coverage differs by dataset, position, and situation; saved or edited models keep their own settings.
Reports use the configuration in Setup and solve each flop independently. You do not need to Build tree or run Solve first just to generate a report.
- Standard report sizes replaces Setup's bet menu with a shared report menu. Turn it off to use your Setup sizes.
- Vs modeled villain applies one available player model from the Preflop export. It does not copy Browse's manual locks or lock both players. The current selection prefers OOP when both players have models.
- Open in Browse rebuilds and solves the selected report spot; it does not load a stored full solve. Match the board, action line, sizes, and modeling assumptions when comparing frequencies. Different convergence targets can still produce differences.
Use Group by → High card in Chart or Table, click column headings to sort, or switch to Hand classes and Textures. Hover a strategy bar for every action's frequency, including zeroes. Search and delete completed reports from the library.
Screenshots captured from the running app on 10 September 2026, using saved studies.
Use the app's Save controls before closing or restarting the server. Files stay in the repository's saves/ directory:
| Folder | Contents |
|---|---|
saves/ |
Postflop solve sessions |
saves/preflop/ |
Preflop game sessions |
saves/profiles/ |
Saved player models |
saves/reports/ |
Reports and their per-node summaries |
To update, finish or stop running jobs, save both sessions, and pull the latest code with git pull. Backend changes require a server restart and rebuild; reopening an already-running Windows server keeps its existing binary. See Windows updates and shortcuts for the full procedure.
- Postflop solves the configured heads-up tree. Results depend on the ranges, sizes, rake, locks, and convergence target you choose.
- Preflop uses an approximate continuation model. New games use coupled-deck equity for pots with three or more players; heads-up pots retain the configured realization model. It does not solve the full multiway game through the river, and overlapping tight ranges can still expose large card-removal errors. Model details.
- Player models combine observations and inference. Sparse or unsupported situations need estimates; a dataset label is not proof that every displayed hand frequency was observed.
- Postflop is heads-up only. There is no ICM, and saved solves use GTOpen's own format.
Older preflop saves keep their original equity model. Re-solve continues that model; save your work, then Build game → Solve to use the updated model in a fresh game.
| Topic | Guide |
|---|---|
| Model coverage | Ignition · CoinPoker · Player types |
| Model behavior | Evidence labels · Raise sizing · Contextual preflop · Contextual postflop |
| Running studies | Overnight reports · Research setup |
| Development | Engine, CLI, API, and runtime settings · GPU benchmarks · Preflop performance results · Preflop research |
The code is organized into crates/solver (Rust engine), crates/server (HTTP API), and web (browser UI).
# CPU tests
cargo test --release -p solver
# GPU integration tests (requires CUDA)
cargo test --release --features gpu --test gpu --test preflop_gpu -- --test-threads=1