PaLM 2
Last reviewed
Sources
10 citations
Review status
Source-backed
Revision
v2 · 1,369 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
10 citations
Review status
Source-backed
Revision
v2 · 1,369 words
Add missing citations, update stale details, or suggest a clearer explanation.
PaLM 2 is the large language model Google announced on May 10, 2023, at its I/O developer conference as the successor to the original PaLM (Pathways Language Model). Google released it in four sizes named Gecko, Otter, Bison, and Unicorn, used it to power Bard and more than 25 products, and built specialized variants on top of it for medicine (Med-PaLM 2) and cybersecurity (Sec-PaLM). It served as Google's flagship general-purpose model for roughly seven months before being superseded by the Gemini family in December 2023. Google framed PaLM 2 as a model with "better multilingual and reasoning capabilities" that was "more compute-efficient than its predecessor PaLM," with "faster and more efficient inference."[1][2][3]
Google announced PaLM 2 during the opening keynote of Google I/O on May 10, 2023. It was the model meant to answer competitive pressure from OpenAI, whose ChatGPT had launched in late 2022 and whose GPT-4 had arrived in March 2023, and Google was widely seen as racing to respond. PaLM 2 made its public debut packaged with a long list of products it would immediately improve, and Bard moved onto it the same day.[2]
The first PaLM was a 540-billion-parameter model that Google introduced in 2022 to demonstrate its Pathways training infrastructure. It was large, capable, and expensive to run. PaLM 2 took a different tack. Rather than chasing raw scale, Google described it as a model with "better multilingual and reasoning capabilities" that was "more compute-efficient" than its predecessor, with "faster and more efficient inference."[3] In the words of the technical report, "This improved efficiency enables broader deployment while also allowing the model to respond faster, for a more natural pace of interaction."[3] The shift reflected a broader industry move in 2023 away from ever-larger parameter counts and toward models that were cheaper to deploy while performing as well or better on downstream tasks.
Google made PaLM 2 available in four sizes, named from smallest to largest after animals: Gecko, Otter, Bison, and Unicorn. The naming let Google match a model to a use case without committing to public parameter figures.[1][4]
| Size | Position | Notable point |
|---|---|---|
| Gecko | Smallest | Light enough to run on mobile devices, fast enough for interactive on-device applications even when offline; reported to process about 20 tokens per second on a phone |
| Otter | Small to medium | Mid-tier option below Bison |
| Bison | Medium to large | Offered through the PaLM API for text and chat |
| Unicorn | Largest | The most capable variant in the family |
Of these, Gecko drew the most attention. Google said it was "so lightweight that it can work on mobile devices and is fast enough for great interactive applications on-device, even when offline," a claim that pointed toward future on-phone assistants, and reporting at launch put its on-device throughput at roughly 20 tokens per second.[1][5] Bison was the size most developers actually touched, since it backed the PaLM API offerings for text and chat that Google exposed through its cloud platform.[6]
PaLM 2 is a "Transformer-based model trained using a mixture of objectives," according to its technical report.[3] Google emphasized three areas of improvement over the original PaLM.
Multilingual ability was the headline. Google said PaLM 2 "is more heavily trained on multilingual text, spanning more than 100 languages," which "significantly improved its ability to understand, generate and translate nuanced text," including idioms, poems, and riddles. Google reported that the model passed advanced language proficiency exams at the "mastery" level.[1]
Reasoning was the second pillar. Because the training corpus included scientific papers and web pages containing mathematical expressions, Google said the model "demonstrates improved capabilities in logic, common sense reasoning, and mathematics." The technical report backed this with "large improvements over PaLM on BIG-Bench and other reasoning tasks."[1][3] On several BIG-Bench Hard tasks, including multi-step arithmetic, temporal-sequence reasoning, and hierarchical reasoning with Dyck languages, the report said PaLM 2 "improves over PaLM by more than 100%, demonstrating new emerging abilities."[3]
Coding was the third. PaLM 2 was pre-trained on a large body of publicly available source code and was proficient in more than 20 programming languages. Beyond mainstream choices such as Python and JavaScript, Google specifically noted that it "can also generate specialized code in languages like Prolog, Fortran and Verilog."[1]
At I/O, Google said it "announced over 25 new products and features powered by PaLM 2."[1] The most visible was Bard, the company's conversational assistant, which had launched in February 2023 on the older LaMDA model and moved to PaLM 2 at the conference. Google said the switch would give Bard stronger math, reasoning, and coding skills.[2][7] PaLM 2 also fed into Google Workspace features across Gmail, Docs, and Sheets, and into Duet AI, the generative assistant Google was building for Workspace and Google Cloud.[1][6]
For developers, Google released the PaLM API. It was available inside the Vertex AI platform with what Google called enterprise-grade privacy, security, and governance, and also offered as a standalone endpoint, initially limited to a Trusted Testers program.[6]
Two domain-specific models were built on the PaLM 2 lineage:
Google was deliberately vague about how big PaLM 2 was and what it had been trained on. The technical report described architecture and benchmark gains but did not publish parameter counts or the full composition of the training data.[3] The animal-themed size names reinforced this reticence, giving customers a rough ladder of capability without hard numbers.
A week after the launch, CNBC reported figures it said came from internal Google documentation. According to that reporting, PaLM 2 was trained on 3.6 trillion tokens and had 340 billion parameters, compared with 780 billion tokens and 540 billion parameters for the 2022 PaLM. The numbers, if accurate, meant PaLM 2 was trained on nearly five times more text yet was substantially smaller than its predecessor, consistent with Google's public framing of a more compute-efficient model.[4] Google did not confirm the figures, and they should be treated as a press report rather than an official disclosure.
PaLM 2's run as Google's lead model was short. At the same I/O event, Google previewed Gemini, calling it "our next model created from the ground up to be multimodal."[1] Gemini was formally announced on December 6, 2023, and an updated Bard adopted Gemini Pro in its place, ending Bard's roughly seven-month stint on PaLM 2.[10] Within twelve months of its February 2023 launch, Bard had cycled through three backbones: LaMDA, then PaLM 2, then Gemini 1.0. PaLM 2 remained available through the PaLM API for a period afterward, but Google's product and research focus moved firmly to Gemini.