Model comparison
Compare EmbeddingGemma 2 and Gemini Embedding 2 using the same provider-sourced embeddings & vector search rubric. No mystery score and no invented benchmark ranking.
Set your usage. Your estimate updates as you type.
Assumes 600 tokens per page, processed separately. Actual token counts vary. This covers embedding only, not storage, search, or generated answers.
Estimates exclude taxes, tools, cache storage/writes, free allowances and custom discounts. Image estimates cover output only, not prompt or reference-image charges. Quality modes differ by model. Unlisted settings are not treated as free.
| Model | Access | Estimated total (USD) |
|---|---|---|
| Gemini Embedding 2Google | $0.0012 | |
| EmbeddingGemma 2Google DeepMind | Google DeepMind | No reviewed rate |
Quick take
Google DeepMind's open-weight embedding model for local search across text, code, images, audio, and video.
Open weights do not make compute free. Rebuild and evaluate the index when changing models, use the documented numerical precision, and verify truncated vectors on your own data.
Google's multimodal embedding model for placing text, images, video, audio, and PDFs in one searchable vector space.
Media inputs have separate limits and prices, so text-only cost estimates do not describe a multimodal index. Free-tier and paid Gemini API data-use terms also differ.
Compare the published facts
Values use each provider's own published units and limits. A blank means the provider did not publish a directly comparable value in the sources reviewed.
| Embeddings & vector search | EmbeddingGemma 2 | Gemini Embedding 2 |
|---|---|---|
| Embedding priceCurrent provider price per million input tokens or the closest published billing unit. | Open weights; self-hosted | Text $0.20 / 1M tokens; multimodal rates vary |
| Input capacityMaximum content accepted in one embedding input, using the provider's documented token basis. | 8,192 shared tokens across text and media | 8,192 text tokens; media has separate limits |
| Vector dimensionsSupported output sizes; smaller vectors reduce storage while larger vectors may preserve more information. | 768 default; 512, 256, or 128 with re-normalization | 128–3,072; 768, 1,536, or 3,072 recommended |
| Accepted inputsText, code, image, audio, video, PDF, or document-aware input supported by the endpoint. | Text, code, images, audio, and sampled video frames | Text, image, video, audio, PDF |
| Retrieval controlsQuery/document modes, task types, truncation, chunking, or other controls that shape vectors for retrieval. | Task prompts, selective encoders, vector truncation, and normalization | Task types, output dimension, title for retrieval documents |
| Where it runsDirect API, cloud marketplace, private deployment, or self-hosted route documented by the provider. | Local or self-hosted; Apache 2.0 weights | Gemini Developer API and Google AI Studio |
How to choose
Start with the job you need to complete, then validate cost, access, and policy details on your exact provider route.
Inference can run locally without sending source content to an embedding API. Data retention and network access depend on the application and services you connect.
Google's Gemini API terms distinguish unpaid and paid services: content from unpaid services may be used to improve products, while paid-service prompts and responses are not used to improve products.
Provider and API links
Frequently asked questions
EmbeddingGemma 2: Google DeepMind's open-weight embedding model for local search across text, code, images, audio, and video. Gemini Embedding 2: Google's multimodal embedding model for placing text, images, video, audio, and PDFs in one searchable vector space.
Consider EmbeddingGemma 2 when your priority is Local and offline multimodal search. Consider Gemini Embedding 2 when your priority is Multimodal search across text and media. Test both with your own data and provider route before committing.
No. This comparison aligns provider-published facts for the Embeddings & vector search category. It does not claim a universal winner or combine incompatible third-party benchmark scores.
