PaLM 2

Claims and evidence

• Released 10 May 2023Announced at Google I/O and offered the same day in Bard, through PaLM API sign-ups and in Vertex AI. ● Primary source[1]
• Retired 21 April 2025Vertex AI models text-bison (PaLM 2 for Text) and chat-bison (PaLM 2 for Chat). ● Primary source[3]Retired models: text-bison, chat-bison [4]first paragraph [5]first paragraph

• Status RetiredRetirement as listed for text-bison and chat-bison on Vertex AI (now Gemini Enterprise Agent Platform); archived Vertex AI references name these models PaLM 2 for Text and PaLM 2 for Chat. No current Google page gives the shutdown date of the PaLM API. ● Primary source[3]Retired models: text-bison, chat-bison [4]first paragraph [5]first paragraph
• Replaced by Gemini 2.5 Flash-LiteRecommended upgrade for text-bison and chat-bison as listed when checked on 2026-10-01. ● Primary source[3]Retired models: text-bison, chat-bison [4]first paragraph

• Successor of PaLM ● Primary source[1]Introducing PaLM 2 [2]Abstract

• Change · Languages Pre-trained much more heavily on multilingual text spanning more than 100 languages; the data mix also adds more code, mathematics and parallel multilingual documents.Compared with PaLM ● Primary source[1] [2]
• Change · Architecture Trained with a tuned mixture of pre-training objectives, following UL2; Google says this makes it more compute-efficient than PaLM, with faster inference.Compared with PaLM ● Primary source[2]
• Change · Size Google says the biggest PaLM 2 model is significantly smaller than the biggest PaLM model, although it used more training compute; exact sizes were not disclosed.Compared with PaLM ● Primary source[2]
• Change · Availability Offered in Google products such as Bard and to developers through the PaLM API and Vertex AI, whereas PaLM was a research model.Compared with PaLM ● Primary source[1] [6]Table 30 (PaLM Model Card): Application

• Input text ● Primary source[2]Table 31 (PaLM 2 Model Card): Input(s)
• Output text ● Primary source[2]Table 31 (PaLM 2 Model Card): Output(s)
• Feature MultilingualTrained more heavily on multilingual text spanning more than 100 languages. ● Primary source[1]Introducing PaLM 2
• Parameters Not disclosedGoogle withheld model sizes; the report says only that the largest PaLM 2 model is significantly smaller than the largest PaLM model. ● Primary source[2]section 1; Table 31 (PaLM 2 Model Card): Model Architecture, Model Stats
• Access API, consumer appPaLM API sign-ups for developers, Vertex AI for Google Cloud customers, and Bard. ● Primary source[1]Powering over 25 Google products and features

Lineage

Variants

Related AI Radar coverage

AI Radar coverage starts in June 2026; no coverage linked yet.

All model releases from Google on AI Radar →