A local heads-up poker research bot

Two release lines play two-player no-limit hold'em, with no rake or ante and fresh stacks each hand. Both play tabular, linearly weighted, opponent-sampled CFR averages.

v0.5.2, heads-up 100 BB (newest).

• The model: a 4B-node average (seed 2026100601), the same training run as v0.5.0 and v0.5.1 continued to 4B nodes, with public-history translation for off-menu bet sizes.
• Head to head: it beats v0.5.1 by +17.90 [14.98, 20.83] BB/100. Two independent seeds repeat the gain over their own 2B models: +18.7 and +19.4 (#236).
• Scripted opponents (translation on), BB/100:

No opponent shows a severe regression from v0.5.1 on the same deals; loose-aggressive improves by +35.

• Limits: no external benchmark has been played, and pot-size pressure still beats it.
• Model card · install and run

v0.4.2, heads-up 20 BB.

• The model: a 10B-node average, fixed seed 2026100601. It is the default on the 20 BB table; v0.4.0 and v0.4.1 remain selectable.
• Results: the three retained 10B lineages beat matched-seed 1B v0.4.1 directly by +3.50 [+1.63, +5.37] BB/100, and all declared candidate checks pass.

• What the intervals cover: they're nominal paired 95% intervals, conditional on the saved lineages. No full-game exploitability certificate or human or professional strength follows.
• Updated decision table and limits · v0.4.2 model card · earlier v0.4.0 → v0.4.1 comparison

Watch two bots

In the local web app, choose Bot-vs-bot spectator, select two pinned models, then use Play, Pause or Step. Each decision shows its exact policy probabilities, selected action, lookup status and acting bot's legal perspective. Sessions retain both release/model/manifest identities and replayable hand history; v0.4.0 and v0.4.1 remain selectable. See the spectator guide and independent audit command.

Quick start

Use Python 3.11, Git and the GitHub CLI. From a fresh checkout:

git clone https://github.com/dberweger2017/deepcfr-texas-no-limit-holdem-6-players.git
cd deepcfr-texas-no-limit-holdem-6-players
python3.11 -m venv .venv
source .venv/bin/activate
python -m pip install -r requirements-play.txt
mkdir models
gh release download v0.4.2 --pattern 'O10B-HU20-opponent-sampled-average-seed-2026100601.jsonl.gz' --dir models
gh release download v0.4.1 --pattern 'O1B-HU20-opponent-sampled-average-seed-2026100601.jsonl.gz' --dir models
gh release download v0.4.0 --pattern 'B100M-HU20-current-seed-2026093001.json.gz' --dir models
python -m scripts.verify_v042_model models/O10B-HU20-opponent-sampled-average-seed-2026100601.jsonl.gz
python -m scripts.verify_v041_model models/O1B-HU20-opponent-sampled-average-seed-2026100601.jsonl.gz
python -m scripts.verify_v04_model models/B100M-HU20-current-seed-2026093001.json.gz
python -m src.play_api.server --models-dir models

mkdir models and downloads fail rather than overwrite existing artifacts. If the directory exists, inspect it or use a fresh destination. Installation builds the pinned native pokers engine; follow its build error instructions if your platform needs a Rust toolchain.

Open http://127.0.0.1:8765/ and enter the token from results/play-web/access.token locally. Keep tokens out of URLs and screenshots. Choose v0.4.2 (default), v0.4.1 or v0.4.0, choose bet sizing, create a session and deal. Seats alternate each hand. Model choice is frozen for that session; refresh resumes it, including after server restart. Use a new session to change models. All three pinned files are required for model selection; a missing or corrupt default causes an error, never a silent replacement. The average readers use compact storage; the M4 runtime/replay smoke peaked at 2.76 GiB with all three releases. Allow at least 3 GiB of RAM.

Restricted mode uses the concrete training menu. Free sizing executes any native-legal integer-chip human wager exactly; the bot keeps its trained menu and unchanged uniform fallback on missing/zero-mass keys. This does not solve off-tree strategy. Play/replay guide, human benchmark guide.

| Model | Default / selection | Training and inference | | --- | --- | --- | | v0.4.2 | Default in this checkout | #185 O, seed 2026100601, 10B nodes, opponent-sampled average; exact hash in model card | | v0.4.1 | Still selectable and downloadable | #165 O, seed 2026100601, 1B nodes, opponent-sampled average; exact hash in model card | | v0.4.0 | Still selectable and downloadable | #116 R1, seed 2026093001, 100M nodes, current regret-matched policy |

For a single-model legacy launch, --policy PATH loads v0.4.0; --o-candidate PATH loads the exact v0.4.1 O1B policy (unchanged legacy option). Use --data-dir to keep separate private journals. The server binds only to loopback; see the SSH tunnel guide.

Limits and release plan

Each release uses fresh roots, predeclared paired gates, a direct incumbent match and independent replay of every action/settlement. Bounded LBR gives a lower bound on exploitability; turn/river weakness is conditional on fixed trees and ranges. Small opponent panels remain imprecise. The older neural Deep CFR work is not a complete Pluribus reproduction, and no six-player or 100-BB strength follows from this preview. Research history, roadmap, release notes.

| Release | Game and focus | | --- | --- | | v0.4.0 | Heads-up 20 BB research preview and local table | | v0.4.1 | Heads-up 20 BB average-policy play; release and verified assets | | v0.4.2 | Heads-up 20 BB 10B-node average; stable release and verified assets | | v0.4.x next | #166 turn-search results, then trainer/storage options and finer abstraction | | v0.5.0 | Heads-up 100 BB 1B-node average with translation; internal and scripted evaluation only, as no suitable free public 100 BB opponent has been found; release and verified assets | | v0.5.1 | Heads-up 100 BB 2B-node average with translation; beats v0.5.0 head to head by +29.5 BB/100; release and verified assets | | v0.5.2 | Heads-up 100 BB 4B-node average with translation; beats v0.5.1 head to head by +17.9 BB/100, repeated on two other seeds; release and verified assets | | v0.5.5 | Heads-up 200 BB, trained at that depth and benchmarked against Slumbot; results and uncertainty published | | v0.6 / v0.7 | Three players / four and five players | | v0.8 | Six players, 100 BB; confirmed profit against the scripted pool | | v0.9 / v1.0 | Changing lineups/stacks and full-table play / lower-end professional benchmark |

What 1.0 means

A credible professional reference, predeclared multi-seed confirmation, legal information-safe play and reproducible evidence. Later milestones keep their own acceptance checks.

Developers and licenses

Repository map and commands, artifact storage, and maintenance audit.

The play-only installation excludes neural research dependencies. For the full research and test environment, install requirements-dev.txt, then run python -m pytest -q. CI also checks observations, sessions, arenas, solvers and recovery. Read AGENTS.md, rules and the observation contract before changing behavior.

Repository code is MIT. The pinned pokers upstream has no verified license grant in its repository/package metadata; the owner licenses their changes to that fork under MIT, while the original authors' terms remain unresolved. License audit. Release assets contain inference only, no engine binary or private journal; reports and referenced papers retain their own rights.