Gemini Nano 2
Google DeepMind · 2023-12-06
Gemini Nano 2 is the larger of Google's two Gemini 1.0 Nano models, engineered for on-device deployment and targeting high-memory devices. Google reports a 3.25B-parameter size, distillation from larger Gemini models, and 4-bit quantization for deployment; because the size is rounded, no exact parameter count is recorded. The technical report documents a 32,768-token context and evaluates Nano 2 on both language and pixel-input tasks.
Benchmark scores
| Benchmark | Score |
|---|---|
| BBH (BIG-Bench Hard) | 42.4 |
| BoolQ | 79.3 |
| Capability | 95.3 |
| MATH | 22.8 |
| MathVista (mini) | 30.6 |
| MMLU | 55.8 |
| MMMU | 32.6 |