Gemini Embedding 2
Gemini Embedding 2 is Google's multimodal embedding model, which maps text, image, video, audio and PDF inputs into a single unified embedding space. It takes up to 8,192 input tokens and returns Matryoshka-style flexible output dimensions from 128 to 3072 (768, 1536 and 3072 recommended), covers over 100 languages for text, and expresses task types as instructions written into the prompt rather than as an API parameter.
Strengths
- Maps text, image, video, audio and PDF into a single unified embedding space, enabling cross-modal semantic search
- Flexible Matryoshka output dimensions from 128 to 3072, with truncated vectors automatically normalized
- 8,192-token input limit, four times the 2,048 of the earlier gemini-embedding model
- Over 100 languages supported for text embeddings
- Marked Stable and generally available on the Gemini API, not a preview endpoint
- Free of charge on the Gemini API free tier
Caveats
- The embedding spaces of gemini-embedding-001 and gemini-embedding-2 are incompatible, so upgrading requires re-embedding every existing document
- The 8,192-token limit is shared across all modalities in a request, and per-request media caps are tight: at most 6 images, 180 seconds of audio, 120 seconds of video, and 1 PDF of at most 6 pages
- Task types must be written as text instructions inside the prompt (for example "task: search result | query: {content}") rather than passed as an API parameter, which makes prompt formatting part of retrieval quality
- Paid-tier text input costs $0.20 per 1M tokens, a third more than the $0.15 charged for gemini-embedding
- Proprietary and API-only, with no published benchmark scores on the model page, so no benchmark rows are recorded here
- Vendor
- Context Window
- 8,192 tokens
- License
- Proprietary
- Release Date
- 2026-04-22
- Modalities
- textimageaudiovideo
Capabilities
Vision
Audio
Video
Tool Use
Resources
Pricing
Free of charge on the Gemini API free tier. On the paid tier, text input is $0.20 per 1M tokens, versus $0.15 per 1M tokens for the earlier gemini-embedding model.