Gemini 1.5 Pro
Mixture-of-experts Gemini 1.5 Pro matched 1.0 Ultra with less compute and offered a 1M-token context (10M tested in research).
Sparse MoE architecture with a standard 128K window and a private-preview 1M window, enough for an hour of video or 700,000+ words. 99% needle-in-a-haystack recall at 1M tokens; beat 1.0 Pro on 87% of benchmarks. The Gemini 1.5 report is arXiv 2403.05530 (2024-03-08).
- Date
- Thursday, 15 February 2024
- Lab
- Google DeepMind
- Kind
- model
- Access
- closed API
Figures
| Measure | Value | Measured by |
|---|---|---|
| Needle-in-a-haystack recall at 1M tokens | 99% near-perfect to 10M tokens in the technical report | company |
| Benchmarks beating Gemini 1.0 Pro | 87% comparable to 1.0 Ultra | company |
Limited preview via AI Studio and Vertex AI. Public API preview 2024-04-09; GA 2024-05-23 (Gemini API changelog).
Sources
- blog.google/technology/ai/google-gemini-next-generation-model-february-2024/
- arxiv.org/abs/2403.05530
This record was checked against its sources on 6 October 2026. How we check