Claims and evidence
- Announced Google Research blog post; the first arXiv version of the paper is dated 2022-04-05. Primary source[1]
- ReleasedRelease date unknown
- Status Research only Primary source[2]Appendix E, Table 30 (PaLM Model Card): Known Limitations
- Output text Primary source[2]Table 30 (PaLM Model Card): Output(s)
- Context window 2048 tokensTraining sequence length, used for all PaLM models. Primary source[2]section 5, Sequence length
- Parameters 540BPaLM 540B, the model presented in the announcement; Table 1 lists 540.35 billion parameters. The 8B and 62B models are variants. Primary source[1] [2]Table 1
- Access research only Primary source[2]Table 30 (PaLM Model Card): Application; Known Limitations
Lineage
Variants
PaLM 8B
Same dates as PaLM.
- Variant Inline variant in this record. Primary source[2]section 2.1, Table 1
Differs in:
- Parameters: 8B Evidence not assessed [2]Table 1
PaLM 62B
Same dates as PaLM.
- Variant Inline variant in this record. Primary source[2]section 2.1, Table 1
Differs in:
- Parameters: 62B Evidence not assessed [2]Table 1
PaLM 540B
Same dates as PaLM.
Related AI Radar coverage
AI Radar coverage starts in June 2026; no coverage linked yet.