China’s DeepSeek challenges rivals with high-performance, low-cost V4 AI model

DeepSeek has accelerated its momentum in the artificial intelligence race with the release of its V4 model, a move aimed at solidifying its position among global technology leaders.

The Chinese AI startup announced on Friday the preview release of its long-anticipated V4 large language model. Based in Hangzhou, the company stated that this rollout provides users with an immediate opportunity to personally test the model’s new capabilities and features.

The announcement comes more than a year after the debut of the company’s R1 reasoning model, which caused a significant stir in global technology markets due to its surprising performance and cost efficiency. Consistent with previous iterations, DeepSeek’s latest upgrade is open-source, allowing developers to download the code, execute it on local systems, and, in many instances, modify it.

Available in two versions tailored to size requirements—”pro” and “flash”—DeepSeek reported that V4 demonstrates robust performance against domestic rivals, particularly in agent-based tasks, information processing, and reasoning workflows.

“DeepSeek’s V4 preview is a major show of strength,” Neil Shah, Vice President of Research at Counterpoint Research, told CNBC. He noted that the model offers lower inference costs compared to previous versions. In the context of AI, inference costs refer to the computational and financial resources required for a trained model to generate an output.

DeepSeek further highlighted that the V4 model has been optimized for use with popular agentic tools, such as Anthropic’s Claude Code and OpenClaw. Wei Sun, Senior AI Analyst at Counterpoint, emphasized that V4’s benchmarking profile suggests it can deliver “excellent agent capabilities at a significantly lower cost.”

Founded in 2023, DeepSeek first garnered international attention in late 2024 with its free and open-source V3 model. At the time, the company claimed the model was trained at a fraction of the cost of its competitors using less powerful chips.

The R1 model, released in January 2025, sparked concern among investors by matching or exceeding the results of the world’s leading large language models. The revelation that R1 was built in just two months with a budget of less than $6 million—using lower-capacity Nvidia chips—raised questions regarding US leadership in AI and the massive infrastructure spending of Western tech giants.

However, Ivan Su, Senior Equity Analyst at Morningstar, suggested that V4 is not expected to create the same level of market shock as R1. Su noted that investors have already priced in the reality that Chinese AI is both competitive and inexpensive.

Su added that DeepSeek’s latest move places other Chinese open-source models in a position of direct competition. “This is a framework that did not exist during the R1 era, and that alone shows how much domestic competition has intensified,” Su said. Since the launch of R1, DeepSeek has faced increased competition from new models released by industry giants such as Alibaba and ByteDance.

The question of which chips were utilized to train and support the V4 model remains a central focus. Chinese tech giant Huawei confirmed on Friday that its newest computing cluster, powered by Ascend AI processors, is capable of supporting DeepSeek’s V4 model.

Despite this confirmation, the extent to which Huawei chips were used in training compared to American-made Nvidia hardware remains unclear. Due to evolving export controls implemented by Washington, Chinese developers are restricted from directly purchasing Nvidia’s most advanced AI semiconductors.

During this period, Beijing has reportedly intensified efforts to develop its domestic chip industry, encouraging technology firms to adopt alternatives from local manufacturers like Huawei.

Wei Sun of Counterpoint noted that V4’s ability to run natively on domestic silicon could have major implications, potentially strengthening Beijing’s AI sovereignty and further reducing reliance on Nvidia.

Following DeepSeek’s announcement, shares of Chinese semiconductor manufacturers traded in Hong Kong saw significant gains, with SMIC rising 9% and Hua Hong Semiconductor climbing 15%. Conversely, shares of other AI players, including MiniMax and Zhipu, recorded declines.