Orient: DeepSeek Harness update accelerates application deployment, domestic computing power demand surges

Orient issued a research report stating that the DeepSeek V4.1 Flash model has only 552B parameters and achieves superior results on multiple benchmarks with fewer parameters. Compared with the previous-generation model, its demand for HBM is reduced to 1/4, and its demand for SSD is reduced to 1/8. The bank expects this to advance the penetration rate of domestic AI large models and accelerate the expansion of domestic AI commercialization scenarios. The bank believes that the iteration of domestic AI large models will drive a surge in demand for domestic computing power and switch chips. Related tickers: Cambricon (688256.SH, not rated), Hygon Information Technology (688041.SH, Buy), Suzhou Centec Communications (688702.SH, not rated). Orient's main views are as follows: DeepSeek V4.1 Flash is faster and stronger, accelerating the realization of inclusive AI According to DeepSeek's official website, the DeepSeek V4.1 Flash model has only 552B parameters, with input/output activated parameters of 8/16B respectively. At the same time, it adopts a new pretraining method and larger-scale reinforcement learning post-training. In multiple benchmark tests, V4.1 Flash achieves superior results with fewer parameters, surpassing large models such as Kimi K3, GLM-5.3, and GPT5.6-Sol; moreover, V4.1 Flash effectively reduces the KV cache size. Compared with the previous-generation model, its demand for HBM is reduced to 1/4, and its demand for SSD is reduced to 1/8. Currently, the peak-hour pricing of DeepSeek V4.1 Flash is 2/8 yuan per million tokens for input (cache miss)/output. The bank believes that DeepSeek V4.1 Flash has made significant progress in both performance and inference cost optimization, and is expected to advance the penetration rate of domestic AI large models and accelerate the expansion of domestic AI commercialization scenarios. DeepSeek Harness completes multiple updates, accelerating the AI commercialization process The DeepSeek V4.1 Flash model has undergone specialized training and optimization for DeepSeek Harness: (1) Enriching file processing and preview methods. The web interface now supports uploading various types of files including images and PDFs, and the model can read them on demand through file tools; (2) Optimizing long-session performance and user experience. This update optimizes the loading, recovery, and continuous dialogue performance of long sessions, reduces memory usage, and optimizes the storage format; (3) Making sub-Agent communication and task control more flexible. The parent Agent and sub-Agents that can continue dialogue support two-way communication, allowing information to be supplemented and task direction to be adjusted during execution; (4) Adding the experimental feature Agent Teams. Agent Teams allows the main Agent to create multiple team members and split, assign, and track work through a shared task list. Members can send messages to each other. The bank believes that the DeepSeek Harness update lowers the user threshold and optimizes the user experience, while Agent teams is expected to effectively raise the upper limit of AI Agent task capabilities and may accelerate the expansion of AI commercial scenarios. Upgraded performance of domestic frontier large models drives a surge in demand for domestic computing power Domestic AI large models are accelerating their iteration. The DeepSeek V4.1 Flash model has significantly improved performance while also offering high cost-effectiveness. The bank believes that the iteration of domestic AI large models will drive a surge in demand for domestic computing power and switch chips. Risk warnings The pace of technological iteration of domestic AI large models is slower than expected; AI commercialization progress is slower than expected; raw material supply is slower than expected; product price reduction risk; chip manufacturing yield improvement is slower than expected.