Startup DeepSeek disrupts the AI industry
Company highlights
DeepSeek, a quant company, was founded in December 2023 by Liang Wenfeng.
- On November 2, 2023, DeepSeek Coder models were released.
- On December, 2024, DeepSeek-V3 was released, a versatile and cost-effective large language model.
- On January 20, 2025, DeepSeek-R1 and DeepSeek-R1-Zero were released, DeepSeek鈥檚 most capable models for complex reasoning and problem solving. In addition, six smaller 鈥渄istilled鈥� versions enable operation on local devices. DeepSeek AI Assistant, a chatbot leveraging DeepSeek-V3, was also made available.
- On January 28, 2025, DeepSeek Janus-Pro, a set of multimodal models for image generation, was released.
DeepSeek model highlights
- DeepSeek-R1 is the first open-source model to match the performance of OpenAI鈥檚 o1 across a range of core tasks.
- DeepSeek models are open source. Its weights have been published and they have all been made commercially available, enabling organizations to build with this technology without any licensing constraints.
- DeepSeek introduces new reinforcement learning techniques to train the model, drastically reducing the time and complexity for developing the model.
AI customers and companies are questioning if they are overpaying for AI capabilities
DeepSeek model highlights
According to DeepSeek researchers, it cost US$6 million to train its chain of thought model and took only two months to build, using older and slower NVIDIA H800 chip technology (DeepSeek API Docs). In contrast, in 2024, OpenAI secured US$6.6 billion to pursue artificial general intelligence with its chain of thought model.
DeepSeek costs vs. OpenAI costs (Liu et al., 2024)
DeepSeek R1 API (1 million tokens) |
OpenAI o1 API (1 million tokens) |
| 55 cents for input | $15 for input (over 27x higher costs than DeepSeek) |
| $2.19 for output | $60 for output (over 27x higher costs than DeepSeek) |
Resource utilization comparison
DeepSeek achieved results with 2.78 million GPU hours (DeepSeek API Docs), significantly lower than Meta鈥檚 30.8 million GPU hours for similar-scale models.
Market reaction to the introduction of DeepSeek has been unprecedented
China鈥檚 Cheap, Open AI Model DeepSeek Thrills Scientists
Nature, January 23, 2025
DeepSeek Forces a Global Technology Reckoning
The New York Times, January 27, 2025
Nasdaq Falls 3% as Tech Stocks Dive on DeepSeek AI Fears
Barron鈥檚, January 27, 2025
Why DeepSeek鈥檚 AI Model Just Became the Top-Rated App in the U.S.
Scientific American, January 27, 2025
China鈥檚 DeepSeek AI Dethrones ChatGPT on App Store
CNBC, January 27, 2025
DeepSeek-R1 Is a Boon for Enterprises 鈥� Making AI Apps Cheaper, Easier to 91制片厂, and More Innovative
VentureBeat, January 27, 2025
The Rise of DeepSeek: 20 Times Cheaper With the Same Results
CTech, January 27, 2025
Biggest Market Loss in History: Nvidia Stock Sheds Nearly $600 Billion As DeepSeek Shakes AI Darling
Fortune, January 27, 2025
China's Chatbot Juggernaut DeepSeek Could Mark the End of US Supremacy in AI
Australian Broadcasting Corporation, January 28, 2025
Why DeepSeek Prompted a $1 Trillion Tech Sell-Off
Business Insider, January 28, 2025
Will China鈥檚 Open-Source AI End U.S. Supremacy in the Field?
Washington Post, January 28, 2025
Alibaba Releases AI Model It Says Surpasses DeepSeek
Reuters, January 29, 2025
Alibaba Unveils Qwen 2.5 AI Model, Says It Outperforms GPT-4o and DeepSeek-V3
MSN, January 29, 2025
DeepSeek performance benchmarks rival OpenAI o1
(Source: Guo et al., 2025)
Note: The performance benchmarks above need to be validated by a third party.
DeepSeek-R1-Zero is the first large-scale model trained solely using reinforcement learning
DeepSeek-V3-Base |
A base frontier model with 671 billion parameters. |
Group Relative Policy Optimization (GRPO)
|
This is the core reinforcement learning algorithm used to train R1-Zero. The model generates multiple responses for a given prompt and the responses are then compared within the group. This process eliminates the need for human-annotated data or human intervention. |
DeepSeek-R1-Zero |
First open-source model trained solely with large-scale reinforcement learning (RL) instead of supervised fine-tuning (SFT) as an initial step. However, a few challenges were uncovered, including poor readability and language mixing. |
DeepSeek R1-Zero Release Date: January 20, 2025
(Source: Vellum, 2025)
DeepSeek R1 uses GRPO to address the challenges of training AI models using reinforcement learning
- Initializes the training process using an initial set of solutions (cold start data), which is often generated by a more advanced model (fixing readability challenges).
- Apply GRPO process, as performed to produce R1-Zero, to enhance reasoning capabilities.
- Evaluate data/solutions based on criteria such as correctness, coherence, and relevance. Highest quality solutions proceed to the next step.
- The DeepSeek V3 Base model is used to generate high-quality chain-of-thought examples. These examples then serve as the training data for the SFT process.
- A second stage of GRPO is applied, building upon the fine-tuning with new data, to further improve reasoning capabilities across a wide variety of scenarios (e.g. helpfulness, safety).
(Source: Vellum, 2025)
Carefully assess the use of any open-source LLM before implementation
Strengths
- Reduce costs
- Improve AI performance
- Leverage new innovative techniques to reduce the time and complexity to train the model
Weaknesses
- Privacy concerns
- Potential bias
- Content filtering
Opportunities
- Deploy more advanced AI models at lower costs
- Greater deployment options (cloud, on premises, hybrid)
- Reduce costs
Threats
- Current AI investments may be disrupted
- Possible cybersecurity/jailbreaking threats
- Future innovations will continue to disrupt the AI ecosystem
Prepare for accelerated change and disruption
DeepSeek is a disruptive technology that will change how models will be trained and deployed.
- It demonstrates the viability of open-source vs. proprietary models.
- It will drive widespread testing and adoption for the correct use cases.
- New innovative ways to train the model will accelerate development.
- The AI ecosystem of vendors will be assessing and integrating new capabilities as a result.
- Pricing changes to a variety of AI offerings are likely.
- Expect integration and adoption of new AI training techniques to accelerate.
Action items to consider for your organization
- Be aware of the model鈥檚 limits and restrictions regarding privacy and openness.
- Ask your AI vendor on future pricing for their offerings in light of DeepSeek offerings.
- Prepare for more innovations, especially regarding AI software methodologies.
- Days after the DeepSeek announcement, Alibaba announced its Qwen 2.5 AI model, claiming it can outperform GPT-4o and DeepSeek-V3.
Consider Testing:
- DeepSeek-V3: For flexibility and multitasking; ideal for multilingual, multidomain applications
- DeepSeek Coder: For software development, debugging, and IDE integration
- DeepSeek-R1: For those use cases involving complex logical tasks
- DeepSeek Janus-Pro: For those use cases requiring multimodal responses with images
Related Info-Tech research and other resources
Prepare to Negotiate Your Generative AI Vendor Contract
91制片厂 Your AI Solution Selection Criteria
Adopt a Structured Acquisition Process to Ensure Excellence in Gen AI Outcomes
DeepSeek Resources
Works cited
Guo, Daya, et al. 鈥淒eepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning.鈥� arXiv, arXiv:2501.12948, 22 Jan. 2025. Web.
Kirkovska, Anita. 鈥淏reaking down the DeepSeek-R1 training process鈥攏o PhD required.鈥� Vellum, 24 Jan. 2025. Web.
Liu, Aixin, et al. 鈥淒eepSeek-V3 Technical Report.鈥� arXiv, arXiv:2412.19437, 27 Dec. 2024. Web.
鈥淢odels & Pricing.鈥� DeepSeek API Docs, n.d. Web.