We are pleased to announce the addition of DeepSeek-R1 to Snowflake Cortex AI in its preliminary phase. This model, developed by DeepSeek and trained through large-scale reinforcement learning (RL) without supervised fine-tuning (SFT), offers performance comparable to that of OpenAI-o1 in key areas such as mathematics, code generation, and reasoning. According to performance tests published by DeepSeek, DeepSeek-R1 ranks as a leader among open-source models and competes with some of the most advanced closed-source solutions on the market.
Customers can now apply for early access to DeepSeek-R1 in Cortex AI. During this private preliminary phase, our focus will be on ensuring an experience that aligns with our principles of ease, efficiency, and trust.
The model will be available in a serverless mode, allowing for both batch (SQL function) and interactive (Python and REST APIAn API, short for “Application Programming Interface”) inferences. To request access to the preview, interested parties should contact their sales team. The model’s availability will be limited to the requested account.
The model is hosted in the United States, within Snowflake’s service perimeter. We do not share data with the model provider.
Once the model is generally available, customers will be able to manage access through role-based access control (RBAC). Account administrators will be able to restrict access by selecting approved models according to governance policies.
Snowflake Cortex AI
Snowflake Cortex AI is a set of integrated features and services that include fully managed LLM inference, fine-tuning, and RAG (Generative Augmented Retrieval) for structured and unstructured data, enabling customers to quickly analyze unstructured data alongside their structured data and accelerate the creation of AI applications. Customers can access industry-leading LLMs, both open-source and proprietary, and easily integrate them into their workflows and applications. Snowflake has embraced the open-source ecosystem with support for multiple LLMs from Meta, Mistral, and Snowflake itself. We believe that this open access and collaboration will pave the way for accelerated innovation in this space.
DeepSeek-R1
According to the DeepSeek post on GitHub, they directly applied reinforcement learning to the base model without relying on supervised fine-tuning as a preliminary step. This approach allowed the model to explore the chain of thought (CoT) to solve complex problems, resulting in the development of DeepSeek-R1-Zero. They also mention that the initial model demonstrated capabilities such as self-verification, reflection, and generating long CoTs, but faced challenges like endless repetition, poor readability, and mixing languages. To address these issues, the DeepSeek team describes how they incorporated cold-start data before RL to improve reasoning performance.
The team implemented low-precision FP8 training and a load balancingLoad balancing is a strategy used to distribute workloads strategy with no auxiliary loss, leading to state-of-the-art performance with significantly reduced training computational costs.
Using DeepSeek-R1 in Cortex AI
With Snowflake Cortex AI, accessing large language models is straightforward. There is no need to manage integrations or API keys. Governance controls can be consistently implemented across data and AI. You can access the models in one of the supported regions. Additionally, you can access from other regions with cross-region inference enabled. You can enable Cortex Guard to filter potentially inappropriate or unsafe responses. Security barriers reinforce governance by applying aligned policies to filter harmful content.
What’s Next?
According to DeepSeek, this is the first open-source model that demonstrates that the reasoning capabilities of LLMs can be incentivized solely through RL, without the need for SFT. Cortex AI provides easy integration via SQL functions and REST API, and Cortex Guard allows customers to implement the necessary security controls. The Snowflake AI research team plans to enhance DeepSeek-R1 to further reduce inference costs. Customers can achieve cost-performance efficiencies with DeepSeek-R1 and accelerate the delivery of generative AI applications. This advancement paves the way for future breakthroughs in this area.