DeepSeek models are available for use as managed APIs andself-deployed models on Vertex AI. You can stream your responses toreduce the end-user latency perception. A streamed response usesserver-sentevents (SSE) to incrementally stream the response.

Managed DeepSeek models

Note: The DeepSeek R1 model is not a Google product, and itsavailability in Vertex AI is subject to the terms for "SeparateOfferings" in the AI/ML Services section of theService SpecificTerms, and separate terms foundin the relevant model card.

DeepSeek models offer fully managed and serverless models as APIs. Touse a DeepSeek model on Vertex AI, send a request directlyto the Vertex AI API endpoint. When using DeepSeek models as amanaged API, there's no need to provision or manage infrastructure.

The following models are available from DeepSeek to use inVertex AI. To access a DeepSeek model, go to itsModel Garden model card.

DeepSeek-OCR

DeepSeek-OCR is a comprehensive Optical Character Recognition (OCR)model that analyzes and understands complex documents. It excels atchallenging OCR tasks, including recognizing mathematical formulas andprocessing text that is curved, rotated, or overlapping.

Go to the DeepSeek-OCR model card

DeepSeek-V3.2

DeepSeek-V3.2 is a model that harmonizes high computationalefficiency with superior reasoning and agent performance. DeepSeek'sapproach is built upon three key technical breakthroughs: DeepSeek SparseAttention (DSA), scalable reinforcement learning framework, and large scaleagentic task synthesis pipeline.

Go to the DeepSeek-V3.2 model card

DeepSeek-V3.1

DeepSeek-V3.1 is a hybrid model that supports both thinking mode andnon-thinking mode. Compared to the previous version, this upgrade bringsimprovements in hybrid thinking modes, tool calling, and thinkingefficiency.

Go to the DeepSeek-V3.1 model card

DeepSeek R1 (0528)

DeepSeek R1 (0528) is the latest version of the DeepSeek R1 model.Compared to DeepSeek-R1, it has significantly improved depth of reasoningand inference capabilities. DeepSeek R1 (0528) excels in wide range tasks,such as creative writing, general question answering, editing, andsummarization.

Considerations

For production-ready safety, integrate DeepSeek R1 (0528) withModel Armor, which screens LLM prompts and responses for various security and safetyrisks.

Go to the DeepSeek R1 (0528) model card

Use DeepSeek models

For managed models, you can use curl commands to send requests to the Vertex AI endpoint using the following model names:

• For DeepSeek-OCR, usedeepseek-ocr-maas
• For DeepSeek-V3.2, usedeepseek-v3.2-maas
• For DeepSeek-V3.1, usedeepseek-v3.1-maas
• For DeepSeek R1 (0528), usedeepseek-r1-0528-maas

To learn how to make streaming and non-streaming calls to DeepSeek models, seeCall open model APIs.

To use a self-deployed Vertex AI model:

1. Navigate to theModel Garden console.
2. Find the relevant Vertex AI model.
3. ClickEnable and complete the provided form to get the necessary commercial use licenses.

For more information about deploying and using partner models, see Deploy a partner model and make prediction requests.

DeepSeek model region availability

DeepSeek models are available in the following regions:

Model	Regions
DeepSeek-OCR	us-central1 Max output: 8,192 Context length: 8,192
DeepSeek-V3.2	global Max output: 65,536 Context length: 163,840
DeepSeek-V3.1	us-central1 Max output: 32,768 Context length: 163,840
DeepSeek R1 (0528)	us-central1 Max output: 32,768 Context length: 163,840

What's next

Learn how toCall open model APIs.