Hello everyone, I am Orange Brother! With the rapid development of AI, multiple industries around the world are benefiting from the innovations brought by AI technology. From healthcare, finance, to education and software development, AI has penetrated every aspect of our lives. During the evolution of AI models, the reasoning ability has become a key standard for measuring their effectiveness. Before discussing DeepSeek-R1, I wish everyone a Happy New Year! Feel free to claim a free WeChat red envelope cover, limited quantity, first come first served:
Recently, there has been a very popular and revolutionary new development—DeepSeek-R1. This open-source large language model released by the domestic startup DeepSeek in January 2025 has quickly emerged as a strong competitor to OpenAI’s o1 model due to its efficient reasoning ability and low-cost advantages. Unlike OpenAI’s high-cost and high-hardware requirement model, DeepSeek-R1 has made significant breakthroughs in performance and cost through innovative training methods and open-source strategies.
Core Features and Innovations of DeepSeek-R1
1. Efficient Training Method: Breaking Traditional Limitations
DeepSeek-R1’s training method differs from traditional supervised fine-tuning (SFT); it employs reinforcement learning (RL) for training. This innovative training method not only significantly reduces the training costs of the model but also enables DeepSeek-R1 to autonomously develop more advanced reasoning capabilities. Through reinforcement learning, DeepSeek-R1 can demonstrate exceptional reasoning abilities in various complex tasks, especially in mathematical reasoning and programming.
2. Outstanding Performance: Comparable to Top Models
In several authoritative benchmark tests, DeepSeek-R1 has demonstrated reasoning abilities comparable to, and even surpassing, OpenAI o1. For example, in the AIME 2024 test, DeepSeek-R1 achieved an accuracy rate of 79.8%; in the MATH-500 test, its score reached an astonishing 97.3%. In terms of programming capability, DeepSeek-R1 achieved an Elo rating of 96.3% in the Codeforces test, surpassing most human competitors. These outstanding performances significantly enhance DeepSeek-R1’s competitiveness in the AI field.
3. Open Source and Low Cost: Popularization of AI Technology
Unlike OpenAI’s high API fees, DeepSeek-R1 adopts an open-source strategy and uses the MIT license, allowing global developers to modify and optimize freely. This not only reduces the usage costs for enterprises and developers but also greatly increases the popularity of AI technology. In terms of API charges, DeepSeek-R1’s input cost is only $0.55 per million tokens, and the output cost is $2.19, which is much lower than OpenAI’s similar services, greatly lowering the threshold for using AI technology.
4. Multi-Stage Training Pipeline: Stronger Reasoning Ability
DeepSeek-R1 also adopts a multi-stage training pipeline, combining cold-start data, reinforcement learning, and supervised data. This training strategy enhances the model’s performance in handling complex tasks, especially in scenarios that require deep reasoning, allowing DeepSeek-R1 to provide more accurate and efficient solutions.
Why Is DeepSeek-R1 the New Trend in AI Development?
DeepSeek-R1 is not just a high-performance reasoning model; it is also a paradigm of open-source and low-cost, promoting the popularization of AI technology. This innovative training method and open-source strategy enable DeepSeek-R1 to carve out a niche in the competition with OpenAI o1. By significantly reducing costs and improving reasoning capabilities, DeepSeek-R1 provides global developers with a more competitive choice.
Moreover, the reasoning process of DeepSeek-R1 is transparent, allowing users to clearly understand every step of the model’s reasoning process for more precise adjustments and optimizations. This is a distinct advantage compared to many closed-source AI models.
How to Run DeepSeek-R1 Locally Using Ollama?
Ollama is a powerful tool that allows users to run DeepSeek-R1 on local systems without relying on cloud APIs. The system requirements will vary depending on the model size you choose.
Recommended Hardware Configuration: Larger models (such as 70B, 671B) require high-end GPUs (such as A100, H100) and enterprise-level memory configurations. If resources are limited, it is recommended to choose models ranging from 1.5B to 14B for regular reasoning tasks.
Install Ollama
First, you need to install Ollama. Enter the following command in the command line:
Download and Set Up DeepSeek-R1
After installing Ollama, you can use the following command to download the DeepSeek-R1 model:
You can choose different versions of the model according to your hardware conditions, such as 1.5B, 7B, 8B, 14B, 32B, etc.
Run DeepSeek-R1 Locally
To start interacting with the model, run the following command:
This command allows you to send prompts in real-time and receive responses from the model.
Using DeepSeek-R1 in Python Scripts
You can also integrate DeepSeek-R1 into Python scripts by calling it through Ollama’s API:
This allows for convenient use of DeepSeek-R1 in AI-driven applications.
Conclusion: Moving Towards the Future of AI Technology
DeepSeek-R1 is not just an open-source reasoning model; it symbolizes the popularization of AI technology. Through innovative training methods, excellent reasoning performance, low costs, and open-source advantages, DeepSeek-R1 is driving the penetration of artificial intelligence technology into broader application scenarios. From scientific research to business, from education to software development, it provides global developers with a more competitive, flexible, and transparent AI choice.
「Kuankebang Quantitative Club」is a professional community focusing on the integration of AI and quantitative trading, building a knowledge-sharing platform for practitioners, developers, and fintech investors. The community deeply covers six core areas: grid trading, ETF quantification, US stock high-frequency strategies, cryptocurrency arbitrage, statistical arbitrage, and AI machine learning, developing market prediction models and strategy optimization solutions through machine learning and deep learning technologies, helping members meet the diverse needs of stock, futures, ETF, and cryptocurrency markets. The core user group includes quantitative traders, Python technical developers, and financial innovators. We welcome your participation!
Click to Read the Original Article and JoinKuankebang Quantitative Club