PaLM 2
Google's PaLM 2 is the successor to the original PaLM and represents the next generation of large language models. The model appears to excel in advanced reason...
Last verified:
What is PaLM 2?
PaLM 2 is Google's next-generation large language model with state-of-the-art multilingual, reasoning, and coding capabilities. Built on Google's decade of foundation model research, it is designed to be more compute-efficient than its predecessor while delivering significantly improved quality across downstream tasks. The model powers over 25 Google products and features including Bard, Workspace tools, Med-PaLM 2, and Sec-PaLM, bringing advanced AI capabilities to consumers, developers, and enterprises worldwide.
Key features include heavy training on multilingual text spanning more than 100 languages for nuanced understanding and generation of idioms, poems, and riddles; improved reasoning capabilities in logic, common sense, and mathematics through training on scientific papers and mathematical expressions; and excelling at coding across 20 programming languages from popular ones like Python and JavaScript to specialized languages like Prolog, Fortran, and Verilog. PaLM 2 comes in four sizes—Gecko, Otter, Bison, and Unicorn—allowing deployment for diverse use cases from mobile devices to enterprise systems.
PaLM 2 is for consumers using Google products, developers building AI applications through the PaLM API or Vertex AI, and enterprises needing enterprise-grade privacy, security, and governance. Specialized versions include Med-PaLM 2 for medical knowledge and expert-level performance on medical licensing exams, and Sec-PaLM for cybersecurity analysis to detect malicious scripts. The model passes advanced language proficiency exams at mastery level and achieves state-of-the-art results in medical competency.
The model is available through Google Cloud's Vertex AI platform, the PaLM API (now open to all developers), Firebase, and Colab. It processes 20 tokens per second and Gecko's lightweight design enables offline on-device interactive applications. PaLM 2 forms the foundation for Codey (Google's specialized coding model), powers Duet AI for Google Cloud, and expands Bard to new languages.
PaLM 2 pricing
Pricing model: Free
PaLM 2 is available through Vertex AI with enterprise-grade pricing (not publicly disclosed), the PaLM API (pricing not publicly specified on this page), Firebase, and Colab. Med-PaLM 2 will open to a small group of Cloud customers later in summer for feedback. Sec-PaLM is available through Google Cloud. No free tier or specific paid plan pricing details are provided on this website.
PaLM 2 pros
- State-of-the-art multilingual capabilities across 100+ languages
- Improved reasoning in logic, common sense, and mathematics
- Excels at 20 programming languages including Python, JavaScript, Prolog, Fortran, Verilog
- More compute-efficient than predecessor PaLM
- Four model sizes for flexible deployment: Gecko, Otter, Bison, Unicorn
- Gecko works on mobile devices even offline
- Passes advanced language proficiency exams at mastery level
- State-of-the-art medical competency with Med-PaLM 2
- Expert-level performance on US Medical Licensing Exam questions
- Sec-PaLM detects malicious scripts faster than traditional methods
- Available through Vertex AI with enterprise-grade privacy and security
- Open to all developers via PaLM API
- Also available through Firebase and Colab
- Processes 20 tokens per second
- Powers over 25 Google products including Bard and Workspace
PaLM 2 cons
- No free tier mentioned for public use
- 20 tokens per second not extremely fast
- Mobile deployment timeline not committed
- Med-PaLM 2 only open to small Cloud customer group initially
- Sec-PaLM available only through Google Cloud
- Training data includes public source code only (no private code)
- Gecock offline capability limited to specific devices
- Gemini (next model) still in training with no availability date
Frequently asked questions about PaLM 2
What is PaLM 2?
PaLM 2 is Google's next-generation state-of-the-art language model with improved multilingual, reasoning, and coding capabilities. It is more compute-efficient than its predecessor PaLM while delivering significantly improved quality across downstream tasks.
How many languages does PaLM 2 support?
PaLM 2 is heavily trained on multilingual text spanning more than 100 languages, significantly improving its ability to understand, generate, and translate nuanced text including idioms, poems, and riddles.
What programming languages can PaLM 2 code in?
PaLM 2 excels at 20 programming languages including popular ones like Python and JavaScript, as well as specialized languages like Prolog, Fortran, and Verilog, due to pre-training on large publicly available source code datasets.
What are the different sizes of PaLM 2?
PaLM 2 comes in four sizes from smallest to largest: Gecko (lightweight, works on mobile devices offline), Otter, Bison (most capable), and Unicorn.
How can developers access PaLM 2?
Developers can sign up to use PaLM 2 through the PaLM API (now open to all), use it in Vertex AI with enterprise-grade features, or access it through Firebase and Colab.
What is Med-PaLM 2?
Med-PaLM 2 is trained with medical knowledge by Google's health research teams, can answer questions and summarize insights from dense medical texts, achieves state-of-the-art medical competency, and was the first LLM to perform at expert level on US Medical Licensing Exam-style questions.
What is Sec-PaLM?
Sec-PaLM is a specialized version of PaLM 2 trained on security use cases, available through Google Cloud, that uses AI to analyze and explain potentially malicious script behavior and detect threats faster than traditional methods.
How fast is PaLM 2?
PaLM 2 can process 20 tokens per second, which Google notes may be acceptable for some use cases.
What Google products use PaLM 2?
PaLM 2 powers over 25 Google products including Bard (expanded to new languages with coding updates), Workspace features for Gmail, Google Docs, and Google Sheets, Med-PaLM 2, Sec-PaLM, Duet AI for Google Cloud, and the PaLM API.
What is coming after PaLM 2?
Google is working on Gemini, their next model created from the ground up to be multimodal, highly efficient at tool and API integrations, and built to enable future innovations like memory and planning. Gemini is still in training and will be available at various sizes once fine-tuned and safety-tested.