Curiosity · AI Model
PaLM 2
PaLM 2, announced at Google I/O 2023, was Google's answer to GPT-4 in the pre-Gemini era. It powered Bard, Duet AI for Workspace, and MedPaLM 2, and introduced size-scaled variants (Gecko, Otter, Bison, Unicorn) for different cost/quality trade-offs. Superseded by the Gemini line from late 2023 onward.
Model specs
- Vendor
- Family
- PaLM 2
- Released
- 2023-05
- Context window
- 8,192 tokens
- Modalities
- text
Strengths
- Size-tiered variants (Gecko → Unicorn) for different cost/quality needs
- Multilingual pretraining across 100+ languages
- Stronger reasoning than original PaLM at launch
Limitations
- Deprecated on Vertex AI in favour of the Gemini family
- Short 8K-token context vs. modern million-token Gemini Pro
- No vision, audio, or tool-use capabilities
Use cases
- Legacy Bard / Duet AI reference implementations
- Historical benchmark comparisons
- Google Cloud Vertex AI PaLM endpoints (deprecated)
Benchmarks
| Benchmark | Score | As of |
|---|---|---|
| MMLU | ≈78% | 2023-05 |
| HumanEval | ≈37% | 2023-05 |
Frequently asked questions
What is PaLM 2?
PaLM 2 is Google's 2023 flagship dense decoder LLM and the successor to PaLM. It powered Bard and Duet AI for Workspace before being superseded by the Gemini family.
Is PaLM 2 still available?
PaLM 2 endpoints are deprecated on Google Cloud Vertex AI. Google recommends migrating to Gemini 1.5, 2.0, or 2.5 Pro for equivalent and better workloads.
What were the PaLM 2 variants?
Google shipped PaLM 2 in four sizes — Gecko, Otter, Bison, and Unicorn — covering on-device to frontier cloud deployments at different cost and quality levels.
Sources
- Google — PaLM 2 technical report — accessed 2026-04-20
- Google I/O 2023 — Bard and PaLM 2 — accessed 2026-04-20