Models

← All models

PaLM 540B

Google DeepMind · 2022-04-04

PaLM 540B is the largest variant of Google's original Pathways Language Model, introduced in April 2022 as a dense decoder-only Transformer trained with Pathways across two TPU v4 Pods. Google evaluated it primarily through few-shot prompting across language, multilingual, reasoning, and code tasks. The paper reports the model at 540.35 billion parameters and its training exposure at approximately 780 billion tokens, so Atlas does not treat either rounded disclosure as an exact integer.

Benchmark scores