Snowflake is integrating an early preview of DeepSeek-R1in Cortex AI. Customers can apply to access the model within Snowflake Cortex AI.
Snowflake wants to ride the wave of enthusiasm surrounding DeepSeek-R1. That model rocked the AI world. After all, Chinese researchers developed an advanced LLM without access to sophisticated hardware and at a fraction of the price. According to early benchmark results from, the model performs similarly to OpenAI-o1 in areas such as math, programming and reasoning.
Snowflake calls itself a specialist in AI data cloud. The company not only wants to manage and unlock customer data for value creation, but also offer recent AI capabilities to go with it. This is necessary: no one wants to park their data with a party like Snowflake if that implies that AI applications for that data are unavailable on the platform. You can see the integration of DeepSeek in that context.
Private preview
DeepSeek-R1 is currently available in a private preview for serverless inference via both batch processing (SQL function) and interactive applications (Python and REST API). Access is granted only within the requested Snowflake account. The model is hosted in the U.S. within the Snowflake service environment, with no data shared with DeepSeek. This is possible, since the creators made the model openly available.
read also
Inside DeepSeek-AI stirs minds: how does it work, and what’s stolen?
When the model becomes generally available, customers can manage access through role-based access control (RBAC). Administrators can set which models are approved according to their organization’s governance policies.
Cortex AI
The model can be deployed via Snowflake Cortex AI, allowing integrations with SQL and Python without separate API keys or external management tasks. The Cortex Guard function can filter harmful or unwanted output. In addition, a REST API is available for applications outside Snowflake.
Snowflake emphasizes that Cortex AI makes it easy to implement and manage generative AI models. In the future, the company plans to further improve the efficiency of DeepSeek-R1 to reduce the cost of inference.