DeepSeek’s “Tech New Year’s Gift”
On 20 January, 2025, as the Chinese New Year approached, the field of artificial intelligence (AI) in China received a unique “New Year’s gift” — DeepSeek officially launched its next-generation general-purpose large model, DeepSeek-R1. Developed by the team under the High-Flyer Group, this large model quickly sparked a global reaction with its innovative low-cost architecture, exceptional performance, and open-source philosophy. Shortly after its release, DeepSeek-R1 made waves in the international capital markets and quickly climbed to the top of the Apple App Store and Google Play Store download charts, reaching over 140 countries and regions, showcasing its powerful global appeal (as shown in Figure 1).
Technological Breakthrough: Innovations with Challenges
DeepSeek-R1 has demonstrated “academic excellence” across a variety of tasks, particularly in mathematics, coding, and reasoning, proving itself comparable to OpenAI-o1 (as shown in Figure 2). In the SuperCLUE-Science Chinese large model evaluation, it scored 73.64 points, surpassing other models and becoming the first domestic reasoning model to outperform OpenAI’s top model (as shown in Figure 3). This achievement is backed by several cutting-edge technologies — the combination of the Mixture of Experts (MoE) system and Multi-Layer Attention (MLA) mechanism doubles the efficiency in handling complex tasks, while the introduction of Chain-of-Thought (CoT) and Multi-Task Pretraining (MTP) brings the model’s problem-solving approach closer to that of human experts.
While DeepSeek-R1 is receiving praise in the spotlight, China’s AI industry is facing tough challenges behind the scenes. The development of large AI models in China is being tested by U.S. technology restrictions and limited computing power. In October 2024, the U.S. imposed new chip export controls that restricted access to advanced process chips, directly cutting off some overseas computing resources, creating immense pressure on Chinese AI companies. Despite these obstacles, China’s tech sector is pushing forward with innovation. DeepSeek-R1 is a strong example of this effort. By using a Mixture-of-Experts (MoE) design and an FP8 mixed-precision framework, it manages to perform efficiently even with hardware limitations.
Recently, DeepSeek has faced another issue. According to the Financial Times, OpenAI has accused DeepSeek of using “model distillation”—a technique where AI models learn by analyzing the outputs of other advanced models. Many in the industry see this as similar to students learning from top teachers, but OpenAI’s rules forbid this practice. It is still unclear whether DeepSeek has broken any rules, but this debate over AI training methods has sparked wider discussions about industry regulations. This issue goes beyond DeepSeek and could shape the future of AI, influencing how companies share technology and compete in the global market.
Practical Application: AI-Driven Industrial Ecosystems
Despite challenges like technology restrictions and limited computing power, DeepSeek has not stopped moving forward. Through innovation and resource optimization, it has not only made breakthroughs in China’s AI model development but also rapidly turned its advancements into real-world applications. This demonstrates its strong potential to empower industries and tells the story of “AI+” across industries. For example, Kingsoft Office has integrated DeepSeek-Writer API into its WPS intelligent writing feature, which not only accelerates the text generation process but also significantly improves the quality of output. In the financial sector, the joint development between TRS and DeepSeek has led to the creation of a financial public opinion large model, which has been deployed in institutions like CITIC Securities to help financial organizations monitor market trends and generate professional investment reports in real-time.
DeepSeek’s technology also shows immense potential in the healthcare sector. AI-powered medical image analysis, diagnostic assistance systems, and intelligent health management platforms can all leverage DeepSeek-R1’s robust reasoning capabilities to improve diagnostic efficiency and accuracy. In education, DeepSeek has broad application prospects as well. By offering personalized learning and intelligent tutoring, DeepSeek can tailor educational programs based on students’ learning progress, interests, and comprehension abilities, thereby enhancing learning outcomes.
Representing the cutting edge of AI large models, DeepSeek is driving the digital transformation of various industries. As the technology continues to evolve, the application scenarios for large models will expand, ushering industries into a new stage of intelligence and efficiency.
Hong Kong’s Opportunity: The Application of DeepSeek in AIFT
In recent years, Hong Kong has accelerated its pace of innovation in artificial intelligence, supported by favorable policies and a well-developed technological ecosystem. This has attracted numerous research institutions and enterprises to explore AI applications. Particularly in the fintech sector, AI is gradually transforming traditional data analysis methods, enhancing the precision of market insights. The Laboratory for AI-Powered Financial Technologies (AIFT) is dedicated to exploring AI’s real-world applications in industrial areas, especially in the financial field. AIFT reviews and continuously tracks multiple open-source LLMs and deploys them for research and development.
AIFT deployed DeepSeek-V2 on June 24, 2024, along with other open-source LLMs such as InternLM and Qwen for data processing tasks due to their strong language understanding and processing capabilities. AIFT researchers actively monitor updates for these LLMs and adjust the usage accordingly. After DeepSeek-R1 was launched on 20 January, 2025, AIFT recognized its impressive reasoning abilities and deployed it using SGLang (an inference library) on 24 January, 2025, for internal use.
With the deep reasoning capabilities of DeepSeek-R1, AIFT can apply its expertise in many specialized areas. In particular, AIFT are utilizing DeepSeek to predict the future direction of a stock by providing historical stock data from the past N days. AIFT researchers feed the model with historical stock data and visualized candlestick (K-line) charts covering a stock’s price movement over the past N days. These inputs include open, high, low, close prices (OHLC) and trading volume, allowing DeepSeek-R1 to analyze market trends and predict potential future movements.
For instance, AIFT inputted a 50-day K-line chart into DeepSeek-R1 (as shown in Figure 4), which displayed OHLC data and trading volume for each trading day. The visual representation indicated steady growth in the early phase, followed by significant volatility around Day 25, and then a sharp price surge with a subsequent upward consolidation trend.
In this case, DeepSeek-R1 conducted a comprehensive analysis of the stock data and concluded that the stock had been experiencing strong upward momentum and was likely to continue its bullish trend over the next 25 days. Specifically, DeepSeek-R1 output a step-by-step analysis:
1. Trend Analysis
– Closing Price Trend: The closing price has generally been increasing over the 25 days, starting at 181.74 on Day 1 and ending at 206.68 on Day 25. This indicates a strong upward trend.
– Volatility: The stock has experienced some volatility, with significant price swings (e.g., Day 24’s low of 191.72 and Day 25’s high of 206.69). However, the overall trend remains upward.
2. Volume Analysis
– Volume Trends: Volume has been relatively stable, with occasional spikes (e.g., Day 25’s volume of 172,373,300). High volume on Day 25, coupled with a significant price increase, suggests strong buying interest and potential continuation of the upward trend.
3. Key Technical Indicators
– Moving Averages: The closing prices are consistently above the short-term moving averages (e.g., 5-day or 10-day), which is a bullish signal.
– Support and Resistance: The stock has found support around the 190 level and resistance near 196. On Day 25, it broke through resistance, reaching 206.68, which is a strong bullish signal.
– Momentum: The stock has shown strong momentum, especially on Day 25, with a significant price increase and high volume.
4. Recent Price Action
– Day 25: The stock surged from 193.21 to 206.68, breaking through previous resistance levels. This is a strong bullish signal, especially given the high volume.
5. Potential Risks
– Overbought Conditions: The sharp rise on Day 25 might indicate overbought conditions, which could lead to a short-term pullback.
– Market Sentiment: External factors (e.g., macroeconomic news, earnings reports) could impact the stock’s direction.
6. Prediction
Based on the analysis:
– Upward Direction: The stock is likely to continue its upward trend in the next 25 days, supported by strong momentum, high volume, and a breakout above resistance levels.
– Short-Term Pullback: There may be minor pullbacks or consolidation periods, but the overall trend is expected to remain positive.
DeepSeek-R1’s analysis aligns well with traditional technical analysis, demonstrating its ability to identify key trends, volume shifts, and market momentum—making it a valuable tool for stock market forecasting. Compared to traditional quantitative analysis models, DeepSeek offers two distinct advantages:
- Logical coherence and consistency – DeepSeek-R1 generates the same conclusions when presented with identical data and rarely contradicts itself during deep thinking sessions, unlike other models, and this is crucial for financial tasks.
- Rigorous and Detailed Reasoning – when multiple scenarios are possible, DeepSeek can deliver distinct analyses for each scenario, especially when employing the R1 method. Instead of providing generic conclusions, DeepSeek-R1 breaks down potential market scenarios and explains each one in detail. For example, when analyzing whether a stock might rise or pull back, the model does not simply state “up” or “down”—instead, it lays out different possibilities and their logical justifications.
By integrating DeepSeek, AIFT has significantly enhanced the financial data analysis capabilities. It has streamlined the market trend prediction process while providing institutional investors with more intelligent and reliable decision-making tools. In the ever-changing financial landscape, this enables investors to seize opportunities while mitigating risks. Looking ahead, AIFT will continue to explore new AI-driven applications in fintech, pushing the boundaries of what artificial intelligence can achieve in the financial industry.
References
[1] DeepSeek Team. (2025). DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning.
[2] SuperCLUE. (2025). SuperCLUE-Science Chinese Large Model Science Reasoning Evaluation Report.
[3] U.S. Department of Commerce, Bureau of Industry and Security (BIS). (2024). Foreign-Produced Direct Product Rule Additions, and Refinements to Controls for Advanced Computing and Semiconductor Manufacturing Items.
The work described in this article was supported by InnoHK initiative, The Government of the HKSAR, and Laboratory for AI-Powered Financial Technologies (AIFT).
(AIFT strives but cannot guarantee the accuracy and reliability of the content, and will not be responsible for any loss or damage caused by any inaccuracy or omission.)
