Significant developments are emerging in the AI large model sector. Last Friday, March 6th, the Wind DeepSeek Concept Index experienced a late-session rally, shifting from a minor decline at the open to close up 1.87%. Among the concept stocks, Talkweb Information surged to a limit-up gain, Daily Interactive rose over 5%, and UCloud Technology-W increased by 4.5%.
The AI agent OpenClaw has gained explosive popularity, with more than 10 companies taking the lead in its deployment. The significant gains in Talkweb Information and UCloud Technology-W are likely linked to the recent viral success of the AI Agent OpenClaw (formerly "Clawdbot"). Last Friday, nearly a thousand developers and AI enthusiasts gathered at Tencent's headquarters and, with assistance from Tencent Cloud engineers, completed the cloud installation of OpenClaw. According to Tencent Cloud, the installation process, facilitated by their Lighthouse lightweight cloud service for one-click deployment, takes only about five minutes and is free, significantly lowering the technical barrier.
Public reports indicate that, based on Meituan data, searches related to "OpenClaw" have increased by over three times month-over-month. Meituan, in collaboration with Lenovo BaiYing, has launched a remote deployment service for OpenClaw. Consumers can access the service page directly by searching "Lobster Installation" within the Meituan app. Referred to industry as the "Little Lobster," OpenClaw is not a traditional aquaculture product but a highly popular open-source AI agent known for its automation capabilities, touted as "taking over the computer and freeing your hands."
The Longgang District Artificial Intelligence (Robotics) Office in Shenzhen recently solicited public comments on the "Several Measures to Support the Development of OpenClaw & OPC in Shenzhen's Longgang District (Draft for Comment)." The proposal encourages market-oriented, professional platform carriers to establish "Lobster Service Zones" offering free OpenClaw deployment services, with eligible providers receiving certain subsidies. It also suggests providing support for the development and promotion of OpenClaw-type agent tools, offering subsidies of up to 2 million yuan for contributions of key code to international mainstream communities, development and listing of skill packages related to Longgang's advantageous industries on skill trading platforms, and development of applications integrating embodied intelligent devices, subject to approval.
Dongwu Securities pointed out that the new generation of AI Agents, represented by OpenClaw, has completed a fundamental transition from being "Q&A tools" to "actionable executive assistants" that can deeply embed into business workflows and adapt to multi-scenario applications, bringing disruptive efficiency improvements to areas like investment research and office work. The industry views the rise of OpenClaw as pushing AI into a new era of intelligent agents. In the A-share market, based on publicly disclosed information, over ten companies have already deployed OpenClaw or completed adaptation and integration.
Looking at the change in margin balance from March 5th compared to the end of 2025 among these companies, three saw their margin balance increase by over 30%: UCloud Technology-W, Northsec, and Loongson Technology. Additionally, Shunya International, Talkweb Information, iFlytek, and QingCloud Technology-U all saw margin balance increases exceeding 10%.
UCloud Technology-W stated that the company took the lead in launching a relevant image for deploying OpenClaw in late January, positioning itself as an industry pioneer in achieving cloud deployment. Northsec's self-developed XinYuan MiXin product has completed adaptation and integration with OpenClaw. Loongson Technology announced the successful localized deployment of the popular AI agent OpenClaw on hardware equipment based on the fully domestic Loongson 3B6000M chip.
Furthermore, cybersecurity risks warrant attention. Recently, the Ministry of Industry and Information Technology's cybersecurity threat and vulnerability information sharing platform identified significant security risks in some OpenClaw open-source AI agent instances under default or improper configurations, which could easily lead to network attacks and information leaks. It is recommended that relevant organizations and users thoroughly check public network exposure, permission settings, and credential management when deploying and applying OpenClaw, close unnecessary public network access, improve security mechanisms like identity authentication, access control, data encryption, and security auditing, and continuously monitor official security bulletins and hardening recommendations to mitigate potential cybersecurity risks.
There is also news regarding the DeepSeek large model. According to media reports, DeepSeek is set to release its latest large model, V4, which natively supports image, video, and text generation functionalities. This marks the first major version update since DeepSeek launched the R1 reasoning model in January 2025. The V4 large model is expected to further address the market demand for low-cost, open-source models domestically. Since the release of R1 in early 2025, DeepSeek has established a paradigm for open-source models globally, maintaining a steady update cadence, and has published multiple papers in internationally authoritative journals, sharing significant research findings in areas like large model reasoning.
In September 2025, DeepSeek officially released the V3.2-Exp model, which introduced DeepSeek Sparse Attention (a sparse attention mechanism) based on V3.1-Terminus, conducting exploratory optimization and validation for the efficiency of long-text training and reasoning. In December 2025, the company released DeepSeek-V3.2 and DeepSeek-V3.2-Speciale. DeepSeek-V3.2 aims to balance reasoning capability and output length, making it suitable for daily use, while DeepSeek-V3.2-Speciale aims to push the reasoning capabilities of open-source models to the extreme, exploring the boundaries of model performance.
On January 27th this year, DeepSeek officially released its new-generation document recognition model, DeepSeek-OCR2, which achieved a breakthrough in the OCR field, with its core innovation being the pioneering "Causal Flow" visual reasoning logic. Currently, data from OpenRouter, the world's largest large model API aggregation platform, shows that DeepSeek V3.2 consistently ranks within the top ten for token consumption. It was previously reported that DeepSeek's token consumption reached 3 trillion in February, ranking fourth among large models.
Guosheng Securities expressed optimism for the year regarding the application productivity Agent direction and the pan-entertainment multimodal direction, anticipating rapid growth in token consumption driven by application development and the development opportunities brought by AI expansion into global markets. Zhongyuan Securities noted that DeepSeek's increased openness and transparency demonstrate full confidence in its research results and heighten industry anticipation for the release of DeepSeek V4. The institution also suggested that large model costs are expected to decrease, potentially alleviating the current domestic chip shortage. If DeepSeek V4 remains open-source, its enhanced capabilities will provide stronger support for downstream AI application vendors, benefiting AI application deployment. Furthermore, DeepSeek V4 is expected to bring milestone technological breakthroughs, potentially initiating a new paradigm for large model development and accelerating progress toward AGI. Simultaneously, large models may achieve deep integration with domestic chips, with training processes potentially being partially or fully completed using domestic chips.
DeepSeek concept stocks have overall outperformed the Shanghai Composite Index. The 2026 Government Work Report proposed deepening and expanding "AI+" initiatives, promoting the accelerated adoption of new-generation smart terminals and agents, driving commercial and large-scale AI applications in key industry sectors, and fostering new smart-native business formats and models. From an institutional perspective, as large models enter the era of agent execution, major global economies are accelerating the construction of AI governance systems. The synergy between AI application deployment and policy support is accelerating, pushing large models from "rapid growth" towards "regulated development," with technological paradigm shifts and commercial value realization accelerating simultaneously.
CCID Consulting pointed out that the large model market will continue its explosive growth. In 2025, China's large model market size was approximately 49 billion yuan, is projected to exceed 70 billion yuan in 2026, and is expected to reach 100 billion yuan in 2027, driven primarily by MaaS services, industry-specific models, and agent applications. According to Wind data, as of March 6th, the average year-to-date increase for the 45 A-share DeepSeek concept stocks exceeded 6%, outperforming the Shanghai Composite Index overall. Fourteen concept stocks have seen cumulative gains exceeding 10% year-to-date, with SUPCON Technology, Capital Online, Daily Interactive, and UCloud Technology-W all posting cumulative gains over 34%.
SUPCON Technology has surged over 50% year-to-date. The company adopted and updated its base large model early, leveraging its excellent reasoning capabilities and openness to significantly enhance logical reasoning ability and substantially reduce operational costs, injecting new vitality into industrial intelligence. Daily Interactive has gained nearly 35% year-to-date. The company's GAI Station deeply integrates with and is empowered by large language models like DeepSeek, promoting multi-scenario application deployment. UCloud Technology-W has also risen nearly 35% year-to-date. As early as last year, the company stated it had completed adaptation work for the entire DeepSeek model series, utilizing its domestic and international GPU node resources and the UModelverse model service platform to deliver full-capability open-source services for DeepSeek.
Seven well-performing stocks have been targeted by margin traders. Regarding margin financing, the total margin balance for the aforementioned 45 DeepSeek concept stocks on March 5th was 62.275 billion yuan, an increase of nearly 4% compared to the end of 2025. Half of the margin trading targets saw increased positions, with Kingsoft Office, Daily Interactive, and TALF Science recording the highest increases; their March 5th margin balances increased by over 50% compared to the end of 2025.
Filtering further for performance, only seven stocks met the criteria of having a March 5th margin balance increase of at least 3% compared to end-2025, achieving profitability in 2025 with year-over-year growth (including turning a profit from a loss). In terms of margin balance increase, Kingsoft Office, Paratera Technology, Sangfor Technologies, and Talkweb Information all saw increases exceeding 10% on March 5th.
Kingsoft Office's margin balance on March 5th increased by over 55% compared to end-2025. The company reported a net profit of 1.843 billion yuan for 2025, a 12.03% year-over-year increase. In September 2025, the company stated that WPS AI had integrated the DeepSeek series of models. Paratera Technology's margin balance on March 5th increased by over 45% compared to end-2025. As a leading independent supercomputing service provider and general-purpose supercomputing cloud service provider in China, the company saw its 2025 net profit increase by over 80% year-over-year. Its intelligent computing cloud platform has deployed multiple versions of DeepSeek, including R1, V3, V3.1, and V3.2, as well as other mainstream models like Zhipu AI and Qwen.
Talkweb Information's margin balance on March 5th increased by nearly 15% compared to end-2025; the company returned to profitability in 2025. As early as the beginning of 2025, the company, as a strategic partner of Ascend, stated that its "Zhaohan" series AI servers and related products had completed deep adaptation with the DeepSeek series of large models, building a secure and reliable computing foundation for domestic AI large models. Additionally, iFlytek and Desay SV Automotive saw their March 5th margin balances increase by over 8% compared to end-2025.
Desay SV Automotive has been surveyed by 20 institutions year-to-date. The company has established strategic partnerships with several renowned robotics companies and officially released the AICube, an AI computing base for robotics, applying the automotive-grade redundant design from its assisted driving domain controllers to robotics scenarios.