Google publishes the 540B-parameter PaLM
Google presents a 540B-parameter dense language model trained with Pathways; the initial release is a research disclosure, without public weights or a general-purpose API.
What happened
Event details
Google Research introduced PaLM, a 540B-parameter dense decoder-only Transformer. The team used Pathways to coordinate training across two TPU v4 Pods with 6,144 chips, then evaluated few-shot performance in language understanding and generation, reasoning, multilingual tasks and code generation. The April 2022 release centered on the model and its training research, not a public model service. Google did not release PaLM's weights. Commercial access through the PaLM API and Vertex AI arrived in 2023 as a later availability change.
Assessment
Why it matters
PaLM demonstrated Pathways training for a 540B dense model across multiple TPU Pods and documented how scale affected few-shot, reasoning and coding capabilities.
Availability
Access notes
The initial release provided a blog post and paper, but no public model weights or general-purpose API.