Chinese AI startup DeepSeek has launched its V4.1-Flash model, which is priced at a fraction of a cent per million tokens, putting considerable pressure on competitors such as Anthropic PBC and Z.AI Co. This new model is touted to outperform existing popular platforms like Moonshot’s Kimi K3, despite its significantly lower cost. Following the announcement, shares in MiniMax Group Inc. and Z.ai experienced a sharp decline, plunging over 8% in Hong Kong.
The V4.1-Flash model not only boasts a lower price point but also requires substantially less high-bandwidth memory (HBM) and solid-state drive (SSD) storage. This reduction in hardware demand has rattled memory stock investors, leading to a dip of over 3% in shares of major South Korean memory makers like Samsung Electronics Co. and SK Hynix Inc. The news has added to existing concerns in the tech sector, impacting the nascent rebound these stocks had been experiencing since a steep selloff in July.
DeepSeek claims its V4.1-Flash model offers faster inference and higher throughput compared to its predecessors and certain U.S. competitors. On performance benchmarks, DeepSeek reports that V4.1-Flash surpassed Anthropic's Opus 5 and OpenAI's GPT 5.6-Sol on tests like DeepSWE v1.1, CyberGym, and Automation-Bench, although it lagged in the Terminal Bench 3.0 benchmarks. This aggressive move by DeepSeek is seen as a significant challenge to the market share of U.S. AI firms, as the lower prices and reduced resource requirements are already driving a shift in token consumption towards Chinese models.
In a related development, DeepSeek is also planning a substantial investment in AI infrastructure. The company intends to deploy at least 160,000 of Huawei Technologies Co.’s next-generation Ascend 950DT accelerators at a new, large-scale data center being constructed in Inner Mongolia. This move could establish one of the largest known clusters of Huawei AI chips, further advancing China’s efforts to reduce its reliance on Nvidia Corp. and bolster its domestic AI capabilities.
