As competition in artificial intelligence between the United States and China reaches new levels of intensity, prominent Chinese AI startup DeepSeek rolled out previews of its highly anticipated next-generation V4 model lineup on Friday, marking another major milestone in China’s push to advance its domestic AI ecosystem independent of U.S. technology.
The V4 release comes months after DeepSeek’s specialized R1 reasoning model upended global tech markets earlier this year, with the startup claiming it outperformed comparable U.S.-built models at a far lower cost. R1 quickly became a global symbol of China’s rapid progress closing the AI gap with the United States, and expectations for V4 have been building among developers and users eager to compare its capabilities to leading models from U.S. industry leaders OpenAI, Anthropic, and Google. Some analysts initially projected the V4 launch would arrive more than two months ago, to coincide with the Lunar New Year holiday.
DeepSeek’s new V4 family includes two core open-source variants: the high-performance V4 Pro and the lightweight V4 Flash. The startup says the new models deliver sweeping upgrades across three key areas: general knowledge retention, logical reasoning, and agentic functionality — the ability for AI to complete complex, multi-step workflows and tasks without constant human input. One of the most notable shifts in the new lineup is its underlying hardware: unlike prior DeepSeek models that relied on U.S.-made chips from industry leader Nvidia, V4 is powered by chips developed by Chinese tech giant Huawei.
In a statement accompanying the launch, DeepSeek shared internal benchmark results comparing V4 to top U.S. models. The company notes its top-tier V4 Pro Max delivers superior performance on standard reasoning tests compared to OpenAI’s recently released GPT-5.2 and Google’s Gemini 3.0-Pro, though it falls slightly short of OpenAI’s GPT-5.4 and Google’s Gemini 3.1-Pro. The V4 Pro, DeepSeek claims, outperforms Anthropic’s mid-tier Claude Sonnet 4.5 in agentic capabilities and comes close to matching Anthropic’s flagship Claude Opus 4.5. For everyday simple agent tasks, the more efficient V4 Flash matches the performance of V4 Pro, with reasoning capabilities that nearly equal its higher-end counterpart, per the company’s testing.
The V4 launch came just hours after OpenAI introduced its own newest model, GPT-5.5, in a clear sign of the breakneck pace of competition in the global AI race. Both V4 variants also include a game-changing 1 million token context window — the measure of how much text and data an AI can process and retain in a single session. That marks an eightfold increase from the 128,000 token window supported by DeepSeek’s previous V3 model, released in late 2024, and enables the new models to handle far larger datasets, long documents, and complex extended conversations more effectively. The startup also emphasized the V4 lineup is designed to run far more efficiently than prior generations.
Unlike closed, proprietary models from leading U.S. AI developers, DeepSeek makes its core technology open source, allowing outside developers to modify, adapt, and build new tools on top of its models. The company also offers a free publicly accessible chatbot for web and mobile users, helping it gain a large global user base. A January report from Microsoft found DeepSeek usage has grown rapidly across many developing nations, especially in regions where Huawei smartphones dominate the consumer market.
Huawei confirmed its compatibility with the new models in a separate statement Friday, noting its Ascend chip platform and supporting ecosystem work seamlessly with DeepSeek V4. Marina Zhang, an associate professor at the University of Technology Sydney, described the V4 rollout as a “pivotal milestone for China’s AI industry”, particularly amid rising global pressure for countries to build self-reliance in critical emerging technologies. She added that the partnership with Huawei demonstrates that a fully functional Chinese AI ecosystem independent of Nvidia’s market dominance is technically achievable, even as U.S.-China technological decoupling continues.
Lian Jye Su, chief analyst at global technology research firm Omdia, concluded that “Based on the benchmark results, it does appear DeepSeek V4 is going to be very competitive against its U.S. rivals.”
Not all industry observers are convinced the V4 represents a transformative leap forward, however. Ivan Su, senior equity analyst at investment research firm Morningstar, argued that while V4 is a solid, capable update to DeepSeek’s product lineup, it does not deliver the same level of groundbreaking innovation that R1 introduced earlier this year. He noted that domestic competition within China’s AI sector has intensified dramatically since R1’s launch, and that independent third-party testing is needed to verify DeepSeek’s own performance claims, since the startup’s internal comparisons cannot yet be confirmed by outside experts.
The V4 launch comes amid ongoing friction between U.S. AI firms and Chinese developers over intellectual property. Earlier this year, Anthropic publicly accused DeepSeek and two other China-based AI labs of running “industrial-scale campaigns” to steal its technology via a process called knowledge distillation, a method that trains a smaller model by feeding it the outputs of a more powerful competitor model to replicate its capabilities. OpenAI made similar accusations in a letter to U.S. lawmakers, and this week Michael Kratsios, chief science and technology adviser to U.S. President Donald Trump, repeated the claims, accusing Chinese tech firms of distilling leading U.S. AI systems to “exploit American expertise and innovation.”
The Chinese embassy in Washington has pushed back against these allegations, framing them as unjustified efforts by the United States to stifle competition from Chinese tech companies.