News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
PaLM 2

PaLM 2

Google's multilingual Transformer language model in four sizes, from on-device to largest

PaLM 2 is a Transformer-based language model from Google, announced in May 2023 as the successor to PaLM. It was trained with a mixture of objectives on multilingual text, scientific papers and source code, and Google says it improves reasoning, multilingual ability and coding while running faster than its predecessor.

Share on WhatsApp

  • Announced May 2023 as the successor to PaLM
  • Trained on text in more than 100 languages
  • Comes in four sizes: Gecko, Otter, Bison and Unicorn
  • Gecko is small enough to run on mobile devices, even offline
  • Powers more than 25 Google products and features, Google says

What it is

PaLM 2 is a large language model built by Google and introduced in May 2023. It follows PaLM and is described as a Transformer trained with a mixture of objectives. The company reports gains in multilingual work, reasoning and programming, plus faster and cheaper inference than the earlier model. It ships in four sizes named Gecko, Otter, Bison and Unicorn. Gecko is small enough to run on a phone, even without a connection. Larger versions sit behind Google products and enterprise services. The technical report stresses that pre-trained models, fine-tuned variants and the products built on them are separate things, so product behaviour may not match reported results.

How it came about

Google announced PaLM 2 at its I/O event in May 2023, presenting it as the next generation after PaLM. The technical report was posted to arXiv on 17 May 2023 and later revised, with a third version in September 2023, listing 128 authors led by Rohan Anil and Andrew Dai. At the same event Google said more than 25 products and features would use the model. The company also said its Brain and DeepMind research teams were being merged into one unit, Google DeepMind, and that a successor model, Gemini, was already in training.

Inside the model

The model is a Transformer trained with a mixture of objectives rather than a single recipe. Google says its training mix leans heavily on multilingual text covering more than 100 languages, along with scientific papers and web pages carrying mathematical expressions, and large amounts of publicly available source code. That mix is credited for gains in translation, logic, common sense and programming. The report gives no parameter counts, layer counts, context window or compute figures. It notes that user-facing products add their own pre- and post-processing, and that the underlying models can change over time.

What it can do

Google reports stronger multilingual understanding, generation and translation, including idioms, poems and riddles, and says PaLM 2 passes advanced language proficiency exams at a 'mastery' level. Reasoning is said to improve on BIG-Bench and other logic and maths tasks over PaLM. For code it handles Python and JavaScript well and can also write Prolog, Fortran and Verilog. The report describes stable results on responsible AI evaluations and inference-time control of toxicity without hurting other capabilities. No benchmark numbers are given in these pages.

How to use it

PaLM 2 is offered through the PaLM API, opened to all sign-ups from the I/O announcement after a developer preview that began in March 2023, and through Vertex AI with enterprise privacy, security and governance. It also runs inside Google products including Bard, Gmail, Google Docs and Google Sheets, and Duet AI for Google Cloud. The four sizes, Gecko, Otter, Bison and Unicorn, let customers pick a model to fit their device or workload. Weights are not released; access is through Google's services.

Safety and testing

The technical report states that PaLM 2 shows steady performance across a suite of responsible AI evaluations and allows toxicity to be controlled at inference time, without extra cost or loss of other abilities. Google says it is committed to releasing helpful and responsible AI tools. Med-PaLM 2, a medical variant, was to reach a small group of Cloud customers for feedback to find safe and useful cases. No incidents are described in these pages.

Reception and debate

Google positions PaLM 2 as more capable and more efficient than PaLM, and highlights its use across more than 25 products and features announced at I/O. The company also notes that bigger is not automatically better and that research creativity matters in building good models. These pages carry no outside reviews, criticism, lawsuits or bans.

Why it matters

PaLM 2 became the base for specialised Google models: Med-PaLM 2 for medicine, which Google says was the first large language model to reach 'expert' level on U.S. Medical Licensing Exam-style questions and is being extended to read x-rays and mammograms, and Sec-PaLM for security analysis of malicious scripts via Google Cloud. It also backs Bard's language expansion and Workspace writing features. Google describes it as a step towards Gemini, its next model, designed from the start to be multimodal and good at tool and API use.

Questions people ask about PaLM 2

What is PaLM 2?

A Transformer-based language model from Google, released in May 2023 as the follow-up to PaLM. Google says it is better at multilingual tasks, reasoning and coding, and runs more efficiently than the earlier model. It comes in four sizes and sits behind many Google products and enterprise services.

How big is PaLM 2?

These pages do not give parameter counts, layer numbers or training compute. Google only says the model family spans four sizes, from Gecko, which can run on a phone, up to Unicorn, the largest. The technical report notes that products built on the models may behave differently from the raw models.

Can I download PaLM 2?

No weights are published. Access is through the PaLM API, which Google opened to developers from the May 2023 announcement after a preview that started that March, and through Vertex AI for enterprise use. It also runs inside Google products such as Bard, Gmail and Google Docs.

What is special about the smaller versions?

Gecko is light enough to work on mobile devices and fast enough for interactive use, including offline. That lets developers fine-tune PaLM 2 for products that cannot rely on a data centre. Google presents the range of sizes as a way to fit many different uses.

What comes after PaLM 2?

Google said at the same announcement that Gemini, its next model, was already in training. Gemini is described as multimodal by design and aimed at tool and API integration, with plans for memory and planning, and it is intended to ship in several sizes like PaLM 2.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on PaLM 2 yet, so the facts on this page come from the sources listed above: arxiv.org/abs/2305.10403, blog.google/technology/ai/google-palm-2-ai-large-language-model/. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.