How to choose a model
Compare intended use, resource needs and evaluation conditions.
Tags describe a focus, not a measured result
General models aim to cover a wider range of decks. Specialist models may focus on particular builds or matchups. Lightweight models prioritize a smaller runtime footprint. These are descriptions of intent; supported cards, resource requirements and comparable matches still matter.
Read the test conditions before comparing win rates
Opponent type, both decks, game count, turn order, policy mode and temperature affect results. WindBot mirror matches, regular deck-pool matches and historical-model matches do not form one interchangeable strength measure. A missing report means evaluation has not been provided, not that the model is necessarily weak or strong.
Distinguish deployment from continued training
ONNX is commonly used for deployment inference, while PTH checkpoints support training or continued training in Core. GKG can package models, vocabulary, runtime semantic assets and optional content together. To continue training, check that a suitable checkpoint is supplied; an ONNX file alone does not restore the original training process.
Choose a release that matches your environment
The same model protocol does not make all vocabularies and semantic resources interchangeable. Review the compatibility information and prefer the complete set of companion assets. Interpret reported hardware requirements in the context of the author’s model size, runtime and test hardware.