Gemini Embedding 2

Gemini Embedding 2 is Google's embedding generation model that's ideal for complex retrieval and analytics tasks.

Gemini Embedding 2 accepts multimodal inputs to generate 3072-dimensional vectors. It accepts images, text, documents, audio, and video inputs and semantically maps the generated vectors into a unified semantic space. This lets you perform tasks, such as searching for an image based on a text description.

Gemini Embedding 2 introduces several features to optimize embedding quality and flexibility:

  • Custom task instructions: By specifying task instructions (for example, task:code retrieval or task:search result) optimize the embeddings for the intended relationships and retrieve more accurate results for the specific goal.

  • Adjustable result size: The model generates a 3072-dimensional float vector, by default. However, you can retrieve a smaller dimensional output by specifying the output_dimensionality parameter.

  • Document OCR: Read OCR from document inputs.

  • Audio track extraction: Extract audio tracks from video inputs and interleave them with video frames.

For more information on how to use Gemini Embedding 2, see Get multimodal embeddings.

Pricing

Model ID gemini-embedding-2
Modalities
Text Input only
Image Input only
Audio Input only
Video Input only
Embeddings Output only
Token limits Maximum input tokens 8,192
Maximum output tokens N/A
Output dimensions Up to 3,072 (with MRL support)
Maximum sequence length 8,192 tokens
Consumption options
Technical specifications Text
  • Maximum input tokens: 8,192
  • Maximum number of files per prompt: 1
  • Maximum number of pages per file (for PDF): 6
  • Maximum file size per file: N/A
  • OCR for scanned PDFs: Not used by default
  • Supported MIME types:
    text/plain, application/pdf
Image
  • Maximum images per prompt: 6
  • Maximum file size per file for inline data or direct uploads through the console: No limit
  • Maximum file size per file from Google Cloud Storage: No limit
  • Maximum number of output images per prompt: N/A
  • Supported MIME types:
    image/png, image/jpeg, image/webp, image/bmp, image/heic, image/heif, image/avif
Video
  • Maximum video length (with audio): 80 seconds
  • Maximum video length (without audio): 120 seconds
  • Maximum number of videos per prompt: 1
  • Supported MIME types:
    video/mpeg, video/mp4
Audio
  • Maximum audio length per prompt: 180 seconds
  • Maximum number of audio files per prompt: 1
  • Supported MIME types:
    audio/mp3, audio/wav
Supported regions

Model availability

  • Global: global
  • United States multi-region: us
  • Europe multi-region: eu
Knowledge cutoff date November 2025
Versions
  • gemini-embedding-2
    • Launch stage: GA
    • Release date: April 22, 2026
  • gemini-embedding-2-preview
    • Launch stage: Public preview
    • Release date: March 10, 2026