A new research report from Orient Securities Company Limited highlights that the DeepSeek V4.1 Flash model, with just 552B parameters, performs better across multiple benchmarks than larger competitors, while slashing hardware requirements. The report suggests this leap in efficiency will accelerate AI adoption and expand commercial applications in China.

The firm believes that the rapid iteration of domestic AI large models will drive significant growth in demand for local computing power and switch chips. Key related stocks highlighted include Cambricon Technologies Corporation Limited (688256.SH) without a rating, Hygon Information Technology Co., Ltd. (688041.SH) with a Buy rating, and Suzhou Centec Communications Co., Ltd. (688702.SH) also without a rating.

Faster and Smarter: DeepSeek V4.1 Flash Paves the Way for Affordable AI

According to DeepSeek's official website, the V4.1 Flash model has a total parameter count of 552B, with input and output activation parameters of 8B and 16B respectively. It employs a novel pre-training method combined with more extensive reinforcement learning post-training. In numerous benchmark tests, V4.1 Flash achieves superior results with fewer parameters, surpassing large models like Kimi K3, GLM-5.3, and GPT5.6-Sol. Moreover, by effectively reducing the KV cache size, the model cuts HBM requirements to one-quarter and SSD needs to one-eighth compared to its predecessor. Peak-time pricing for the model is set at 2 yuan per million tokens for input (cache miss) and 8 yuan per million tokens for output. The research team views these advancements in performance and inference cost optimization as key drivers for increasing the penetration rate of domestic AI models and speeding up the expansion of AI commercialization.

DeepSeek Harness Updates Aim to Accelerate the Commercialization of AI

The V4.1 Flash model has been specifically trained and optimized for the DeepSeek Harness platform. The updates include richer file handling and preview methods, with the web interface now supporting multiple file types such as images and PDFs, which the model can read on demand. The update also optimizes long-conversation performance and user experience, improving loading, recovery, and continued dialogue, while reducing memory usage and refining storage formats. Sub-agent communication and task control are made more flexible, allowing for bidirectional communication between parent and child agents to supplement information and adjust task direction during execution. An experimental feature, Agent Teams, has been introduced, enabling the main agent to create team members who can split, assign, and track work through a shared task list, and communicate with each other. These enhancements are expected to lower the entry barrier, improve user experience, and potentially elevate the capabilities of AI agents, thereby speeding up the realization of AI commercial scenarios.

Upgraded Frontier Models Fuel Surge in Domestic Computing Power Demand

The accelerated iteration of domestic AI large models, exemplified by the high-performance and cost-effective V4.1 Flash, is anticipated to drive robust growth in the demand for domestic computing power and switch chips. Orient Securities Company Limited reiterates its positive outlook, citing this technological momentum as a primary catalyst.

Risk Warnings

Potential risks include slower-than-expected technological iteration of domestic AI models, disappointing commercialization progress, supply shortages of raw materials, product price cuts, and yields for chip manufacturing improving at a slower pace than anticipated.