Should your organization switch to DeepSeek?
DeepSeek continues to make headlines as policymakers, investors and Big Tech executives grapple with the implications of its recent model releases. But should organizations with existing LLM applications consider switching? Before providing key considerations, it’s important to note that DeepSeek is not a single model but a family of models specialized for various tasks and infrastructure restrictions. Some notable examples include:
- DeepSeek-R1: A flagship reasoning model that achieves comparable performance to OpenAI’s o1 in several benchmark tests.
- DeepSeek-R1-Zero: A variant of R1 trained solely through reinforcement learning (RL) without supervised fine-tuning (SFT). This variant is primarily of research interest. While its performance is lower than R1, it is notable because it challenges the previous assumption that supervised fine-tuning (SFT) is essential for competitive performance.
- DeepSeek-R1-Distill-Qwen-32B: A lightweight 32 billion parameter variant of R1 that surpasses OpenAI’s o1-mini in many benchmarks.
- DeepSeek-V3: A model focused on immediate responses, similar to OpenAI’s GPT-4o or Anthropic’s Claude models. It shares the same base pre-training as R1 but is post-trained to use output tokens parsimoniously rather than to think through its reasoning out loud. Its training cost-effectiveness (estimated at $6 million USD) has attracted attention, representing a fraction of the cost associated with other leading models.
- Janus-Pro-7B: A novel lightweight multimodal model that offers combined text-to-image generation and visual question-answering capabilities.
Why Is DeepSeek Controversial?
DeepSeek has sparked debate for several reasons:
1. Cost Reduction
DeepSeek significantly undercuts API pricing of the incumbent frontier model providers, with DeepSeek-R1 priced at $2.19 per million tokens compared to OpenAI’s o1 at $60 per million tokens and GPT-4o at $10 per million tokens at the time of writing. The cost-effectiveness of DeepSeek’s training was behind Nvidia’s record one-day market value loss, with investors pricing in the reduction in future chip demand as other model builders adopt their more efficient training techniques.
2. (Almost) Fully Open-Source
Unlike many “open-source” models that provide model weights but withhold architecture details and inference code, DeepSeek provides access to all of these. While its training datasets and training code remain undisclosed, efforts are already underway on Hugging Face to reproduce the model from scratch.
3. Political Bias & Data Concerns
Critics raise concerns about observed pro-CCP biases in DeepSeek’s outputs and data privacy issues. Potential biases are of considerable ethical concern, as LLMs are increasingly used to aid and automate decision-making processes across industries and society. However, risks to data privacy are related to use of DeepSeek’s API, and can be largely mitigated by using self-hosted or trusted independent cloud platform deployments, as these become available (e.g., AWS, Azure).
4. Disrupting US Big Tech Dominance
DeepSeek’s releases call into question the continued dominance of proprietary model providers such as OpenAI and Anthropic that rely on outperforming open source alternatives to sell access to their APIs. In May 2023, a leaked internal memo at Google identified that although these models still held a slight edge, the gap was closing quickly and predicted that it would not last. While these providers retain some impressive multimodal capabilities currently unmatched by DeepSeek (which is a text-only model), future releases should not be ruled out, and this has not escaped the attention of investors.
Opportunities
DeepSeek offers compelling advantages that could offer significant value to businesses considering switching out the LLMs powering their AI applications:
1. Significant Cost Savings
For companies seeking to reduce AI infrastructure costs, DeepSeek offers a massive pricing advantage. Even if its own API is not recommended for sensitive applications, self-hosting or using trusted alternative providers can provide the same model performance at a fraction of the cost of other frontier models.
2. Customizability Through Fine-Tuning
The open sourcing of DeepSeek’s models presents an opportunity for organizations to fine-tune the model for their purposes. Although DeepSeek-R1’s 672B parameters make self-hosting and fine-tuning prohibitively expensive for most organizations, smaller variants like DeepSeek-R1-Distill-Qwen-32B can be deployed and trained with modest hardware requirements. This model surpasses OpenAI o1-mini in several benchmarks and, when fine-tuned on high-quality data, can match or exceed the performance of larger models—while requiring far less computational power.
Fine-tuning the model can also reduce the necessary prompt length to achieve high quality responses, leading to lower token usage and reduced latency. Several large LLM providers offer fine-tuning as a web-based service, but building fine-tuning pipelines independently avoids vendor lock-in, affording greater flexibility to adapt as newer models and fine-tuning methods become available.
3. Privacy
Somewhat ironically, the opportunity to locally self-host a DeepSeek model may be attractive to those concerned about the lack of data privacy in allowing sensitive data to pass through the internet and external servers of closed-source LLM providers. While anonymization approaches exist to mitigate these risks, they preclude certain use cases that rely on the sensitive data itself to produce effective output, such as healthcare chatbots and virtual assistants.
Adoption Risks
Before making the switch, organizations should carefully evaluate the risks associated with DeepSeek:
1. Data Security
If your LLM application processes sensitive data, security should be a priority. Directly using DeepSeek’s own API could pose security and compliance risks, especially for enterprises handling regulated data. DeepSeek’s privacy policy reserves the right to collect and store information including IP addresses, chat history and keystrokes, with no declared data retention or encryption policies. A safer approach would be to access DeepSeek via a trusted cloud platform like AWS or Azure, either through a managed service such as AWS Bedrock or self-hosting the model on instances within an existing cloud environment.
2. Political Bias Implications
All LLMs have some degree of bias, but DeepSeek’s technical report makes no mention of any attempt to mitigate bias in its training strategy. This contrasts with those of models such as Llama, GPT-4 and Claude, which openly discuss biases observed during their development and steps taken to address them. For low-risk applications with restrictive structured outputs such as spam filters or structured data extraction, this may not be consequential. However, for externally-facing applications like AI chatbots, careful implementation with robust guardrails is essential to prevent unintended reputational risks, and DeepSeek is unlikely to be a suitable choice of model.
3. Model Security and Adversarial Vulnerabilities
With 2025 forecast to be the year of agentic AI, LLMs are increasingly augmented with access to external databases and APIs, and the ability to perform autonomous actions. This increases the consequence of adversarial attacks through “jailbreaks”—malicious prompts that manipulate the model’s behavior. In a security evaluation derived from HarmBench, DeepSeek-R1 was found to be exceptionally vulnerable, with a 100% success rate in simulated attacks, compared to 36% for Claude 3.5 Sonnet and 26% for OpenAI o1. Even with additional guardrails, for applications with a non-negligible attack surface, DeepSeek does not appear to be an attractive option.
In Summary
Consider DeepSeek if:
- Your use case prioritizes cost savings and can use cloud platform hosted models rather than DeepSeek’s direct API.
- Your AI applications include internal use cases with constricted outputs and limited exposure to bias concerns and security vulnerabilities.
- You could benefit from fine-tuning a smaller model like DeepSeek-R1-Distill-Qwen-32B to optimize for specific tasks at lower infrastructure costs.
Proceed with caution if:
- Your application is external-facing and vulnerable to adversarial attacks.
- You operate in highly regulated industries with strict data security and compliance requirements.
- Your application has bias-sensitive outputs, such as customer-facing chatbots.
DeepSeek heralds a major shift in the AI industry, increasing competition and choice for organizations. Companies focused on cost efficiency with particular use cases may find it a valuable option—especially when delivered through trusted cloud platforms or self-hosted. However, those use cases requiring robust security, bias mitigation, and strong defenses against adversarial attacks might find existing commercial models a better fit for now. As the AI field evolves, organizations must carefully weigh the trade-offs between cost, performance, security, and bias mitigation to align their LLM strategies with their goals and risk tolerance.
