Google 现已在 Gemini API 和 Vertex AI 中提供 `gemini-embedding-001` 文本模型。该模型始终位于大规模文本嵌入基准评测 (MTEB) 多语言排行榜的榜首，在科学、法律、金融和编码等不同领域，性能优于之前的 Google 模型和外部产品。它支持 100 多种语言，并具有 2048 个最大输入令牌长度。一项关键创新是利用 Matryoshka Representation Learning (MRL，俄罗斯套娃表示学习)，使开发人员能够将输出维度从默认的 3072 扩展，从而优化性能和存储成本。该模型通过免费和付费层提供，价格为每 100 万个输入令牌 0.15 美元，并且与现有的 `embed_content` 端点兼容，未来将支持批量 API。建议开发人员在 2025 年 8 月/2026 年 1 月之前从较旧的实验性/传统模型迁移。
We’re excited to announce that our first Gemini Embedding text model (gemini-embedding-001) is now generally available to developers in the Gemini API and Vertex AI.
This embedding model has consistently held a top spot on the Massive Text Embedding Benchmark (MTEB) Multilingual leaderboard since the experimental launch in March.
Surpassing both our previous text embedding models and external offerings in diverse tasks, from retrieval to classification, gemini-embedding-001 provides a unified cutting edge experience across domains, including science, legal, finance, and coding. Here is how Gemini Embedding compares to other commercially available proprietary models:
More detailed results are available in our technical report*.
Model details
An incredibly versatile model, Gemini Embedding supports over 100 languages and has a 2048 maximum input token length.
It also utilizes the Matryoshka Representation Learning (MRL) technique, which allows developers to scale the output dimensions down from the default 3072. This flexibility enables you to optimize for performance and storage costs to fit your specific needs. For the highest quality results, we recommend using 3072, 1536, or 768 output dimensions.
Rate limits and pricing
We offer both free and paid tiers in the Gemini API, so you can experiment with gemini-embedding-001 at no cost, or ramp up with significantly higher limits for your production needs.
The Gemini Embedding model is priced at $0.15 per 1M input tokens.
Start building with Gemini Embedding
Developers can now access the Gemini Embedding model (gemini-embedding-001) via the Gemini API, which you can start working with for free through Google AI Studio.
It’s compatible with the existing embed_content endpoint.
from google import genai
client = genai.Client()
result = client.models.embed_content(
model="gemini-embedding-001",
contents="What is the meaning of life?"
)
print(result.embeddings)To get started, check out the official developer documentation and cookbooks:
If you are using the experimental gemini-embedding-exp-03-07, you won’t need to re-embed your contents but it will no longer be supported by the Gemini API on August 14, 2025. Legacy models will also be deprecated in the coming months: embedding-001 on August 14, 2025 and text-embedding-004 on January 14, 2026. We highly recommend migrating your projects to our newest model as early as possible.
We can't wait to see how Gemini Embedding unlocks new use cases that weren’t previously possible. In addition, we will have support for Gemini Embedding in the Batch API soon, which enables asynchronous processing of your data for lower costs.
Keep an eye out for future announcements regarding embedding models with even broader modalities and capabilities!
*MTEB benchmark results in the published paper reflect the experimental version of Gemini Embedding, launched in March 2025.
