Huachuang Securities: DeepSeek Wishful Wishful Programming Continues to Decrease AI Application Development Threshold

The Zhitong Finance App learned that Huachuang Securities released a research report saying that on October 17, DeepSeek officially upgraded V4.1 Flash to V4.1-Flash-1017. The bank believes that AI coding is moving from auxiliary code generation to complete project development, which is expected to lower the software creation threshold for non-professional users and promote the evolution of software development from professional programming to natural language driving. AIcoding is expanding to scenarios such as game development, 3D content production, enterprise software, and automated workflows, and platforms with developer portals and application ecosystems are expected to benefit first.

The main views of Huachuang Securities are as follows:

DeepSeek continues to evolve, and wishing-style programming is being implemented at an accelerated pace

DeepSeek V4.1-Flash-1017 enhances learning through self-feedback, focusing on improving complete project development capabilities for game development and 3D modeling scenarios. Four examples, including 3D modeling of the Imperial Palace, the web racing game, the Godot version of “Portal 2”, and Valley of Ages, were officially displayed. The model used Harness to complete a runnable project prototype using Harness. Compared with traditional AIcoding, which focuses on code completion and partial modifications, “wish-style programming” further covers project construction, functional implementation, and interactive design. The bank believes that AIcoding is moving from auxiliary code generation to complete project development, which is expected to lower the software creation threshold for non-professional users and promote the evolution of software development from professional programming to natural language driving.

Programming performance has improved significantly, and domestic models are catching up faster

According to official DeepSeek data, v4.1-Flash1017 achieved a score of 77.0% in the DeepSWE v1.1 test, higher than the previous version of 74.2%, 73.0% of GPT-5.6Sol, and 74.0% of Claude Opus 5; Terminal-Bench 4.0 was raised from 31.2 to 42.3, ProgramBench was raised from 20.3 to 29.1, and Automation-Bench was raised from 54.8 to 60.5, and more Agent and programming indicators have been significantly improved. Although some tests still lag behind leading international closed source models, the new version has shown strong competitiveness in some complex programming tasks. The bank believes that the domestic open source model is catching up with international cutting-edge levels in the AIcoding field, and the focus of model competition will shift from simple code generation to the ability to complete complex engineering tasks from end to end.

Task costs continue to decline, and AIcoding commercialization is expected to accelerate

While improving programming capabilities, DeepSeek's upgrade further shortens task running time and output length, and achieves Pareto improvements in performance and cost. According to the official DeepSWE v1.1 performance cost chart, Xinzhiyuan estimates that the task cost for the new version of the max file is about 1-2 yuan, which is lower than the previous version of 3-4 yuan, while the task pass rate has increased from 74.2% to 77.0%. At the same time, DeepSeek maintains the original API unit price, and adapts and optimizes for DeepSeek Harness, Tencent CodeBuddy, and WorkBuddy. The bank believes that AIcoding competition is shifting from the unit price of tokens to the cost of a single successful task. Performance improvements and cost reductions are expected to jointly drive the penetration of individual developers and enterprise users and accelerate commercialization.

The open source ecosystem continues to improve, and the threshold for AI application development has been further lowered

DeepSeek simultaneously opened the v4.1-Flash-1017 model weight, adopted the MIT open source protocol, and officially opened the self-developed Search/Fetch API. Its underlying search engine has indexed 100 billion global web pages and processed hundreds of millions of search requests per day. Both interfaces are billed for 5 yuan/1,000 successful requests. Combined with agent execution frameworks such as DeepSeek Harness, the company is gradually forming a development system covering basic models, tool calls, task execution, and feedback optimization. The bank believes that the continuous improvement of the open source model and agent tool chain is expected to reduce application development costs, promote the expansion of AIcoding to scenarios such as game development, 3D content production, enterprise software, and automated workflows, and that platforms with developer portals and application ecosystems are expected to benefit first.

Investment advice

As the big model continues to be iterated and upgraded, the application is expected to accelerate commercialization. It is recommended to focus on: domestic computing power: Cambrian, Haiguang Information, Mu Xi shares, Tianshu Zhixin, Bezao Technology, Moor Threads; Big Model: Smart Spectrum, MiniMax; Supernode: Inspur Information, Zhongke Shuguang, Ziguang Co., Ltd.; AI+ Industry: Central Control Technology, Hikvision, Suochen Technology, Dingjie Shuguang, Ziguang; AIInfra: Dameng Data, Haitian Ruisheng; AI4S: Hualan Co., Ltd., Insilicon Intelligence, Jingtai Holdings, Jietai Technology; AI+coding: Zhuoyi Information.

Risk warning: AI commercialization and payment conversions fall short of expectations; the open source model intensifies price competition in the industry.