Gemini Embedding 2
#1 in Embeddings & Vector DBsgoogle · v2 · since Public Preview: 10. März 2026; GA: 22. April 2026 · 24× · last seen Jun 30, 2026
100
Momentum
Gemini Embedding 2 is Google's first natively multimodal embedding model, mapping text, images, video, audio, and PDFs into a single unified vector space. Built on the Gemini architecture, it generates 3,072-dimensional float vectors by default, using Matryoshka Representation Learning (MRL) for flexible output dimensions. It supports over 100 languages and is available as a cloud API via the Gemini API and Vertex AI. It launched as Public Preview on March 10, 2026, and reached General Availability (GA) on April 22, 2026.
Momentum trend
04.04.03.07.
Features
| Deployment (Self-Hosted/Cloud) | Cloud API only (no self-hosting / on-premises). Available via Gemini API and Vertex AI. |
| Throughput/Latency | No officially published latency/throughput figures. Per user report: up to 70% latency reduction vs. multi-model pipelines (customer quote, not an official benchmark). |
| License | Proprietary – Gemini API Additional Terms of Service (Google). Free tier: data used for product improvement; Paid tier: no use for training. |
| Platform | Gemini API (ai.google.dev) & Vertex AI / Gemini Enterprise Agent Platform (Google Cloud). Ecosystem integrations: LangChain, LlamaIndex, Haystack, Weaviate, Qdrant, ChromaDB, Vector Search. |
| Price | Text: $0.20/1M tokens; Image: $0.45/1M tokens; Audio: $6.50/1M tokens; Video: $12.00/1M tokens. Free tier available (rate-limited). Batch API: 50% discount. |
| Protocol Compatibility | REST API (Gemini API embedding endpoint). OpenAI compatibility library supported (Batch API). SDKs: Google Gen AI SDK (Python, Node.js, Java, and others). |
| Release Date | Public preview: March 10, 2026 | GA: April 22, 2026 |
| Supported Models/Providers | Model ID: gemini-embedding-2 (GA) / gemini-embedding-2-preview. Input: text (up to 8,192 tokens), up to 6 images (PNG/JPEG), video up to 120s (MP4/MOV), audio up to 180s (native), PDFs up to 6 pages. Output: 3,072-dim vectors (MRL: scalable to 1,536, 768, 128). |