No-Limit Hold’em Bot

It’s no surprise that PokerBotAI’s flagship NLH engine – called PokerX – was designed specifically for this version of the game – a dedicated model that was trained on hundreds of millions of real hands of Texas hold’em and fine-tuned by experienced professional players for all levels of buy-ins from micros to mid-high. In contrast, PokerBotAI’s other products were developed using a proprietary hybrid approach combining the strengths of both traditional rule-based systems and modern neural-network architecture.

Unlike other PokerBotAI products, the NLH Engine is not just a concept; it is a fully functional product that is ready to ship. Onboarding typically takes days instead of months.

The primary goal of this page is to explain exactly how the NLH Engine works, how it generates decisions, what kind of expected win rates are possible depending on the level of competition and what you need to do to get up and running.

What makes the PokerBotAI NLH engine work?

The NLH Engine uses the same tri-brain architecture as PokerBotAI’s other products. Three components generate decisions that provide a good balance of raw computation power and opponent awareness.

The hand history layer

The hand history layer analyzes and stores information about every opponent the engine has played against to create a detailed picture of their playing style. As the engine continues to grow and collect hands from live online play, most of the regular opponents at any given table will already have some form of profile prior to sitting down to play against the engine. The hand history layer is the component responsible for generating the profiles.

The neural network layer

The neural network is the brain behind the engine. It has been trained on more than 7 billion synthetic and solver generated hands, and more than 300 million real hands taken from live online play since the early 2000s. Training runs on Nvidia Tesla T4 GPUs (16 GB VRAM) – datacenter hardware, not hobbyist setups – with new models being created approximately monthly. Instead of providing a single deterministic answer, the neural network provides a weighted Decision distribution so that the engine can choose to deviate from pure GTO when exploiting an opponent’s weaknesses is worthwhile.

The experts layer

The experts layer contains specialist algorithms that complement the neural net when there are opportunities to gain an advantage through pure algorithmic means – multi-way pots, very deep stacked games, short deck variants, and edge opportunities such as ICM heavy late stage MTTs. When a situation corresponds to a well-known pattern, the experts layer activates and improves upon the Decision generated by the neural network.

For serious NLH players: this is a model that is specifically optimized for NLH, not a generalized poker model that also supports NLH. Each PokerBotAI model is independently trained for NLH, PLO, and OFC and tailored per Stakes and per room. The NLH model is the most mature in the lineup and has the largest training corpus.

How does the engine generate decisions at the table?

There are several key features of how the engine generates decisions that separate them from a generic GTO solver used in isolation.

A GTO base layer with opponent-specific exploitation

GTO is the fundamental Decision-making process in the engine. Equity aware bet sizing and balanced ranges are combined with opponent-specific exploitation derived from each player’s hand history profile at the table. If the player in front of us is folding too many river bets, we widen our bluff frequency on the river. If they are a calling station, we tighten our Decision making to ensure we only play value-only hands.

Varying Decision intervals & bet sizes using probability distributions

We draw Decision intervals from a distribution rather than specifying fixed time windows – varying by street, stack size, and opponent profile. We also draw bet sizes from continuous distributions rather than rounding to discrete buckets. These two mechanisms allow us to play at large scales without creating a detectible behavioral footprint after thousands of hands.

Auto-mode vs. Manual-mode

As mentioned earlier, there are two operation modes: auto-mode and manual-mode. auto-mode: runs the engine autonomously – simply specify your target Stakes, stop loss limits, number of tables and session length. The engine then operates without supervision. This is suitable for multi-table operations and/or creating volume. manual-mode: operates as an advisory system – analyze each situation and recommend a course of action. Player performs execution themselves. Manual-mode is suited for operators wanting an analytical advantage without dealing with the complexities associated with full autonomous operation.

Win rates in reality

PokerX’s NLH engine has produced consistent win rates across published Stakes ranges. The actual win rate achieved will depend on game type (e.g., 6-Max vs. Full ring), pool quality and operational discipline related to maintaining clean accounts etc… But it appears reliable in terms of achieving edge within sufficient volume.

Cash NLH 6-Max

In Cash games at NL10-NL50 6-Max, the engine regularly achieves win rates between 8bb/100-14bb/100 with statistically significant sample sizes (i.e., > 250k+ hands). At NL100-NL400 6-Max, the rate drops to roughly 4-8bb/100 due to increased average opposition quality. At NL500+, edge remains present however variance and table availability become limiting factors – these are professional Stakes even with a strong engine.

Cash NLH full ring

Win rates in full ring games are similar in terms of lower variance but smaller edge – 5-10bb/100 at low/micro Stakes and 3-6bb/100 at higher Stakes. Lower variance and longer hand frequencies contribute to full ring being more attractive for operators seeking steady volume versus peak hourly rate.

Edge at higher Stakes becomes less significant

Stake density of strong regulars increase rapidly above NL1k and the engine’s edge over the pool tightens. While still potentially profitable in select games, stake selection (picking the best table at optimal times) becomes more critical in determining win rate as opposed to purely model strength. Many operators find it easier to achieve higher dollar/hour results scaling multiple mid-Stakes tables rather than attempting to grind individual high-Stakes tables.

Comparing NLH engines from competitors

With respect to other NLH engines on the market, most fall into one category or another: legacy rule-based systems that miss GTO baseline or general solvers wrapped in basic automation that fail to emulate behavioral mimicry. The chart below highlights key differences relative to running NLH at scale.

Getting up & running with your own NLH engine

The Onboarding process for your own NLH engine is measured in days not months. The product is shipping now!

Our typical Onboarding flow is as follows: initial consultation regarding desired Stakes & rooms, determination of applicable engine build for specified rooms (different rooms require differing Timing & sizing tunings), setting up your Windows or Android environment for use, calibration session & finally launch under monitored conditions for the first week. Players with intermediate technical skill can begin running on tables within one week. Players requiring a completely managed setup will work directly with our deployment team to configure your environment & monitor your performance until handoff.

Your standard NLH engine begins at $700 us dollars. Pricing escalates based on room support, account count & optional add-ons like our dedicated MTT module (separate offering) or administrative console for managing multiple accounts.

The NLH Engine is offered as a standard product with standard documentation & standard support unlike our brand-specific exclusives discussed throughout the rest of this website.

Placement & application – correct use case

Use cases include mainstream NLH on supported rooms @ micro thru mid Stakes using either auto-mode for scalability or manual-mode advisory play. This does not represent correct use cases: only MTTs where a dedicated module would be more appropriate; only rooms utilizing PaiWangLuo anonymity stack (Ignition/Bovada – separately covered); or any operator seeking a fully custom development engagement for their niche room without existing support.

If none of those limitations pertain to your requirements, the NLH Engine represents a working product with a clear delivery path. Reach out on Telegram to scope your setup and begin the Onboarding process – we will outline your target Stakes & the proper engine build(s) for those Stakes, as well as provide realistic outcome bands based upon your specific requirements.