MANILA, Philippines (Feb 2025) — The open-source large language model (LLM) DeepSeek has gained significant attention with the recent launch of DeepSeek V3 and DeepSeek R1. Developed by a Chinese tech firm founded in 2023, the models have sparked curiosity for their advanced capabilities and open-source availability. Here are answers to some frequently asked questions (FAQs) about DeepSeek, provided by the Tenable Security Response Team.
What is DeepSeek?
DeepSeek is a large language model (LLM) created by Chinese company DeepSeek, founded by Liang Wenfeng in 2023. Its models use natural language processing to respond to user prompts with human-like text.
What is a Large Language Model (LLM)?
An LLM is a machine-learning model trained on massive amounts of data to generate human-like responses. These models power popular chatbots and virtual assistants by predicting words and phrases based on context.
Why the Growing Interest in DeepSeek?
DeepSeek’s appeal lies in two factors: it’s open-source and reportedly trained on less powerful hardware, suggesting that high-quality models can be built at a lower cost. In January 2025, the company released DeepSeek V3 and DeepSeek R1, which quickly caught the tech community’s attention.
How Do DeepSeek V3 and DeepSeek R1 Differ?
- DeepSeek V3 uses a mixture-of-experts (MoE) approach, activating only the necessary components for each task. It also incorporates Multi-head Latent Attention (MLA) to reduce memory usage during training and inference.
- DeepSeek R1 builds on V3’s architecture with Multi-token Prediction (MTP), allowing it to predict two tokens simultaneously. It also supports chain-of-thought (CoT) reasoning, which makes its responses more transparent by showing its thought process.
DeepSeek claims R1’s performance rivals OpenAI’s o1 model.
Minimum Requirements for Running DeepSeek
DeepSeek R1, with 671 billion parameters, requires multiple high-end GPUs. However, distilled versions with fewer parameters can run on standard consumer hardware. Here’s an overview of model sizes:
| Model | Parameters | Size on Disk |
|---|---|---|
| 1.5b | 1.1 GB | |
| 7b | 4.4 GB | |
| 8b | 4.9 GB | |
| 14b | 9.0 GB | |
| 32b | 22 GB | |
| 70b | 43 GB | |
| 671b | 404 GB |
Laptops with 8-16GB RAM can handle models with up to 7 or 8 billion parameters.
What Makes DeepSeek Unique?
Benchmark tests suggest DeepSeek R1 performs comparably with models from OpenAI, Meta, and Claude. Its open-source nature and detailed chain-of-thought reasoning set it apart from competitors.
What Is a Frontier Model?
Frontier models represent the latest advancements in LLMs, focusing on complex reasoning and problem-solving. DeepSeek R1, alongside OpenAI’s o1 and o3 models, is among the few considered as frontier models.
Is It Safe to Use DeepSeek?
Using DeepSeek offline by running models locally on your device is generally safer than accessing it via the company’s website or mobile apps. DeepSeek’s privacy policy indicates that user data is stored in China, where privacy protections like the EU’s GDPR don’t apply. Users should be mindful of potential risks.
Is DeepSeek Banned Anywhere?
As of February 2025, several countries, including Italy, Taiwan, South Korea, and Australia, have banned or are reviewing DeepSeek. In the U.S., states like Texas, New York, and Virginia have restricted its use on government devices, along with agencies like the Department of Defense and the U.S. Congress.
Tenable Research continues to monitor the safety and security implications of LLMs like DeepSeek. Stay tuned for updates on their official blog.
