DeepSeek, a startup based in Hangzhou, has unveiled its V4 large language model, the most powerful version to date, featuring 1.6 trillion parameters and a 1 million token context window. Launched on Friday, this model marks a significant departure as it is optimized for **Huawei’s** Ascend AI processors instead of traditional **Nvidia** hardware. The release coincides with a report by Reuters revealing that the U.S. State Department has dispatched a diplomatic cable to embassies worldwide, urging them to caution foreign governments regarding alleged intellectual property theft involving DeepSeek and other Chinese AI firms.
The V4 model is available in two variants: the flagship **V4-Pro**, priced at $3.48 per million output tokens, and the smaller **V4-Flash**, which has 284 billion parameters and costs $0.28. In comparison, **OpenAI** charges $30 per million output tokens for its **GPT-5.4**, while **Anthropic** sets its price at $25 for **Claude Opus 4.6**. DeepSeek acknowledges that while V4 “falls marginally short” of these closed-source models by approximately three to six months in development, it surpasses all other open-source competitors in agentic coding and reasoning benchmarks.
Previously, DeepSeek’s V3 model was trained using 2,048 **Nvidia H800 GPUs**, leading to scrutiny regarding potential restrictions on acquiring Nvidia hardware through intermediaries in Singapore. By transitioning to **Ascend** chips, the V4 model circumvents these supply chain issues. Huawei has confirmed that its full Ascend SuperNode product line, including the latest **950 series processors**, is compatible with V4 from day one. DeepSeek indicated that V4-Pro costs could decrease further as Huawei ramps up production of the Ascend 950 in the latter half of this year.
According to Reuters, the U.S. diplomatic cable requested embassy personnel to address “concerns over adversaries’ extraction and distillation” of American AI models, explicitly naming DeepSeek, **Moonshot AI**, and **MiniMax**. Just days prior, the **White House Office of Science and Technology Policy** released a memo accusing Chinese entities of orchestrating “deliberate, industrial-scale campaigns” to distill American frontier AI systems.
These allegations follow claims made by Anthropic in February, when it asserted that DeepSeek, Moonshot, and MiniMax utilized 24,000 fraudulent accounts to engage in 16 million exchanges with its Claude model. **OpenAI** has also accused DeepSeek of similar distillation practices. In response, China’s foreign ministry deemed these accusations “groundless.” DeepSeek has previously stated that its V3 model was trained on naturally occurring data gathered through web crawling, without utilizing synthetic data generated by OpenAI.
The timing of the diplomatic cable and the V4 launch is particularly notable, occurring just weeks before President **Trump** is slated to meet with Chinese President **Xi Jinping** in Beijing for a summit anticipated to cover semiconductor export controls and intellectual property disputes. The growing scrutiny of Chinese AI firms in the context of international relations underscores the ongoing tensions surrounding technology and intellectual property in the global market.
See also
Amkor Stock Soars 7%, Elevating Q1 Earnings Expectations Amid AI Packaging Demand
Japan’s Draft AI IP Code Risks Regulatory Misalignment with US Innovation Goals
Micron and Amazon Positioned for Growth Amid $34B AI Revenue Surge
AI-Generated Content Surges: 1.3 Billion Videos on TikTok Raise Authenticity Concerns
Nasdaq Surges as AI Stocks Rally: Top Picks Include Nvidia, Amazon, and TSMC
