LLM

Why an LLM can play poker

A raw screenshot means little to a language model. What makes the engine work is everything that happens before the prompt: Pokercat’s game-flow engine parses the table and rebuilds the complete action context — positions, effective stacks, the exact bet sequence on every street, pot odds at the decision point. Given that clean, unambiguous description, current-generation LLMs reason about ranges and lines far better than their reputation suggests.

This also explains the engine’s edge in multiway pots: pre-solved GTO libraries are built for heads-up confrontations, while a text description of a 3-way spot is no harder for an LLM than a 2-way one.

Choosing and blending models

In LLM mode (and as the fallback inside Auto mode) you pick which models answer from the provider list.

Element	Brief
Model multi-select	Enable one or several models. Each enabled model answers every decision independently.
Blend weights	When several models are enabled, each gets a weight and the panel shows a combined recommendation alongside the individual answers.
Streaming output	Answers stream in as they are generated, so the advice starts appearing immediately.

Tip

Two mid-sized models with sensible weights often beat one large model: disagreements between them are exactly the spots where you should slow down and think.

Cost and limits

LLM calls are billed by actual model usage, so cost varies with the model you pick and the complexity of the spot — see Credits & Billing for how charging works. As with any generative model, answers are not deterministic: the same spot can get slightly different advice twice. The Data Hub tracks LLM-advised hands as their own engine so you can judge the results by profit, not by impression.