DeepSeek has been making waves online, especially with the release of its DeepSeek V3 model. This model has outperformed several competitors while keeping training costs low. If you use AI tools like ChatGPT, you may have heard of it. But what is DeepSeek, and why is it a big deal?
Unlike many secretive and expensive AI projects, DeepSeek takes an open-source approach—meaning its code and training details are shared with the public. In this blog, we’ll explain what DeepSeek is, how it cuts costs, and what this means for AI development.
What is DeepSeek?
DeepSeek is a Chinese-based AI company founded in 2023. Despite being new, it has already made a big impact in the AI-world with it’s V3 and R1 models (more on these later). The company was started by Liang Wenfeng, who previously used AI in financial trading. His goal was to build powerful AI models while keeping costs low.
One of DeepSeek’s key strengths is its open-source commitment. More specifically, it operates under an MIT License, which allows unrestricted commercial use and full control over data. Unlike some of the other AI powerhouses, which keep their models more private, DeepSeek shares its code and research. This approach encourages collaboration and faster AI advancements.
How DeepSeek Improves Efficiency
A big reason DeepSeek made such a large splash is due to the various claims about how much it cost to develop.
In short, DeepSeek claimed in their V3 Technical Report that it cost $5 million to train their model, but this is a very narrow view compared to the total cost of development, which is likely much higher. The difference in how these costs are reported has led to a misunderstanding of the actual investment required for AI development when comparing DeepSeek to companies like OpenAI
- How DeepSeek AI Models Were Developed to Beats GPT-4 at 96% Less Cost
- ‘Exaggerated and misleading’: Google DeepMind CEO casts doubts on DeepSeek’s AI efficiency claim
In a 5-hour podcast with Lex Fridman, AI experts Dylan Patel and Nathan Lambert explored how DeepSeek achieves its efficiency. Patel, the founder of SemiAnalysis, specializes in analyzing semiconductors, GPUs, CPUs, and AI hardware, while Lambert is a research scientist at the Allen Institute for AI. Here are the key takeaways from their discussion:
- Smarter Processing (Mixture of Experts – MoE): DeepSeek’s models don’t use their full capacity for every task. Instead, they activate only the relevant parts (experts), reducing computational waste and improving speed.
- Better Focus (Multi-Head Latent Attention – MLA): This method allows the model to focus on key parts of the data more efficiently, cutting down on memory use while maintaining high accuracy.
- Optimized Hardware Use: DeepSeek has fine-tuned how its models communicate with GPUs, reducing lag and making computations more efficient.
- Selective Activation: DeepSeek’s approach includes high sparsity, meaning only a small portion of the model is active at any time. This significantly lowers the amount of computing power required.
- Open-Source Collaboration: By sharing its technology, DeepSeek benefits from contributions worldwide, leading to continuous improvements and increased efficiency.
DeepSeek R1 & V3
DeepSeek has two major AI models: DeepSeek R1 and DeepSeek V3. These models are similar to OpenAI’s GPT-4o and GPT-4o. R1 is optimized for reasoning and efficiency. V3 is a general-purpose model designed to be both powerful and cost-effective. The models between these two companies aren’t the exact same, but are comparable enough to draw a few parallels.
- DeepSeek R1 is comparable to OpenAI’s GPT-4o, excelling in reasoning, coding, and complex problem-solving tasks. However, R1 is open-source, meaning anyone can download and modify it, unlike OpenAI’s closed-source approach.
- DeepSeek V3 is more like GPT-4o, offering strong performance for everyday AI applications such as chatbots and content creation, but with a major focus on efficiency and affordability.
While OpenAI’s models are proprietary and require API access, DeepSeek’s open-source approach makes its models accessible to businesses, developers, and researchers who want more control over AI tools.
How DeepSeek’s Open-Source Approach is Changing AI
- Making AI More Accessible: DeepSeek’s decision to release open-weight models with a permissive MIT license removes barriers, enabling individuals and smaller organizations to build and use AI without requiring the vast resources of big tech companies.
- Encouraging Competition & Lowering Costs: DeepSeek’s efficient architecture, including MLA and MoE, has demonstrated strong performance while using fewer resources. This puts pressure on companies like OpenAI to consider more open models and makes AI more affordable for businesses and developers.
- Promoting Transparency & Collaboration: By openly sharing its research and model code, DeepSeek allows for greater scrutiny and innovation. This fosters trust in AI development and accelerates improvements across the entire industry.
Security and Global Impact
DeepSeek’s growth has sparked discussions about AI hardware and geopolitics. The company uses NVIDIA’s H800 chips, a version designed for sale in China under U.S. trade restrictions, for training its AI models. U.S. export controls on advanced AI chips to China have impacted the hardware supply for Chinese AI companies. There are ongoing investigations into whether restricted chips are being accessed through intermediaries in countries like Singapore and Malaysia, though no definitive conclusions have been reached.
Concerns about security and data privacy have arisen. Some governments, including Australia, have banned DeepSeek’s applications on government devices due to perceived security risks. New York state, South Korea, and Taiwan have also taken similar actions, with Australia citing an “unacceptable risk” to government technology as the reason for the ban.
Local Deployment Options
Because DeepSeek is open source, developers and businesses can run its models locally instead of relying on cloud-based solutions and avoid a lot of these security concerns:
- Hugging Face: DeepSeek’s models are hosted on Hugging Face, allowing users to download them and integrate them into applications using popular machine learning frameworks like PyTorch.
- GitHub and Other Tools: DeepSeek’s GitHub repository provides detailed documentation for deploying models locally.
Running DeepSeek locally provides greater control over the model, improves response times, and reduces reliance on external cloud providers.
To Sum it Up
DeepSeek’s approach demonstrates that advanced AI can be built with significantly lower compute costs through smart techniques like Mixture-of-Experts and multi-head latent attention.
While these innovations promise more affordable and accessible AI, they also raise questions about security, reliability, and the ability of established tech giants to adapt.
As the industry watches these developments closely, stakeholders will need to carefully balance the benefits of efficiency and openness with concerns over data privacy and geopolitical tensions. DeepSeek’s arrival is an important reminder that the AI advancement is speeding up.
Sources
- DeepSeek
- Entrepreneur – Who is Liang Wenfeng?
- Meredith Media – Open-Source vs. Closed-Source AI Models
- MIT License – Wikipedia
- Hugging Face – Research Paper
- Geeky Gadgets – The Story of DeepSeek R1
- Business Today – DeepMind CEO’s Response
- YouTube – DeepSeek Video
- Agility PR – DeepSeek and U.S. Sanctions
- CNN – Australia DeepSeek Ban
- Hugging Face – DeepSeek AI
- GitHub – DeepSeek V3
Author’s Note: I use AI in my writing to help with formatting, readability, and fact-checking. I do my best to double check every source and fact, but just like how AI can make mistakes, so can humans. If I missed anything or if something is incorrect, please let me know by emailing me at jmeredithmkt@gmail.com or connect with me on LinkedIn here.
Leave a Reply