China Securities Co., Ltd.: DeepSeek local deployment and global asset allocation portfolio tracking

China Securities Co., Ltd. released a research report stating that the DeepSeek-R1 can be deployed locally through the Ollama tool. Firstly, download the version suitable for Windows system from the Ollama official website and install it, select the appropriate model version (such as RTX 4090 recommended 32B), and run it through the command line. To enhance user experience, Docker + Open WebUI can be used to build an interactive interface, and even integrated into WeChat as an intelligent agent. For AI investment, ordinary devices can only run smaller DeepSeek models (such as 1.5B, 7B) with slow response times; the 32B model is suitable for complex scenarios, while the 671B and 70B models require enterprise-level hardware support. Cloud deployment raises data privacy concerns, and the miniaturization of the DeepSeek model has increased demand for computing power for the development of intelligent assistants by small and medium-sized enterprises and individuals. In terms of global multi-asset allocation, low-risk portfolios have returned 0.86% this year, with excess returns of 0.40%; medium to high-risk portfolios have returned 3.66%, with excess returns of 3.61%. Key points from China Securities Co., Ltd. include: Introduction to DeepSeek: DeepSeek, established in 2023, is a subsidiary of Fantasia Quantitative, an artificial intelligence company based in Hangzhou. It launched the DeepSeek-V3 model (671B parameters) at the end of 2024, outperforming various open-source models and approaching top closed-source models. In January 2025, DeepSeek released the R1 series model (660B parameters), excelling in various tasks, and introduced several small models benchmarking OpenAI's products. DeepSeek significantly improves generation speed through its innovative technology and provides competitive API service pricing. Local deployment method of DeepSeek: Ollama is an open-source tool for efficiently running large language models (LLMs) on personal devices without relying on the cloud. The DeepSeek-R1 model can be locally deployed through Ollama: first, download the Windows version suitable for the system from the Ollama official website and install it. After installation, the system tray will display the Ollama icon. Next, visit the "Models" page to select DeepSeek-R1 and choose the 32B version based on the graphics card configuration (such as 4090 graphics card with 24G memory) and copy the corresponding command to run it. Then, execute the command in the command prompt window to download and run the model (the 32B version is approximately 19GB). To enhance user experience, a graphical interaction interface can be built using Docker + Open WebUI, or DeepSeek-R1 32B can be integrated into WeChat as an intelligent entity for quick response and deep thinking capabilities. Thoughts on AI investment: By engaging with DeepSeek on its website and with DeepSeek-V3, the hardware requirements for deploying various versions of models can be understood. Ordinary laptops and desktops equipped with only a CPU can barely run DeepSeek-R1-1.5B and 7B, but with slow response times and lack of practicality. The Nvidia RTX 4090 can run DeepSeek-R1-32B faster, but performs poorly when processing the 70B version. Small models like 1.5B, 7B, and 14B are suitable for simple WeChat communication scenarios but cannot solve complex problems; the 32B model has deep thinking capabilities and is suitable for serving customers through WeChat. The complete 671B and 70B models require enterprise-level graphics cards such as A100 or H100 and are not suitable for consumer-level hardware. While cloud deployment is feasible, data privacy issues exist. The high performance of DeepSeek-R1 and its open-source miniaturized models are driving small and medium-sized enterprises and individuals to develop intelligent assistants, such as WeChat customer service, significantly increasing the demand for computing power. Global multi-asset strategy portfolio performance: The absolute return of a global multi-asset portfolio at low risk is 0.86% this year, with an excess return of 0.40% compared to the ChinaBond Wealth Index (Total). The absolute return of a global multi-asset portfolio at medium to high risk is 3.66% this year, with an excess return of 3.61% compared to the Wind FOF Index. Risk warnings: 1. High correlation reduces diversification effect: The core idea of the model is to evenly distribute the risk of the investment portfolio among various assets, seeking equal risk contribution from each asset. However, when the correlation between certain assets is high, the covariance terms in the covariance matrix will be large, increasing the total risk contribution of these highly correlated assets to the portfolio. As a result, the overall risk of the investment portfolio will increasingly depend on these highly correlated assets, reducing the diversification effect of the risk-parity model. 2. Changes in market environment may lead to model failure: The effectiveness of quantitative models is based on backtesting historical data, but changes in future market environments may differ significantly from historical data, causing the model to fail. For example, changes in macroeconomic conditions, investor trading behavior, or local gaming may affect the actual performance of factors, thereby making risk parity or maximum diversification strategies unable to achieve the expected results. 3. Limitations of asset selection: The effectiveness of the strategy depends largely on asset selection. The choice of assets and market volatility will have a significant impact on the strategy's performance. Investors need to adjust their strategies flexibly based on market conditions and their risk preferences, and be aware of the risk of model failure.