DeepSeek, the Chinese artificial intelligence startup that rattled global markets last year, unveiled preview versions of its latest flagship model on Friday, marking a new escalation in the intensifying AI rivalry between China and the United States.
The release of DeepSeek V4 comes amid heightened anticipation from developers and analysts eager to assess how the system stacks up against leading U.S. models, including those from OpenAI, Anthropic and Google. The launch also follows criticism from some American firms, which have accused DeepSeek of leveraging their technologies in the development of its own systems.
Originally expected earlier this year around the Lunar New Year, the V4 rollout introduces a suite of open-source models, including “pro” and “flash” variants. The company says these models deliver significant gains in knowledge, reasoning, and so-called “agentic” capabilities — referring to the ability of AI systems to autonomously execute complex, multi-step tasks and workflows.
A notable shift in the new generation is its hardware foundation. DeepSeek confirmed that V4 runs on chips developed by Huawei, signaling a move away from reliance on U.S. suppliers such as Nvidia. The transition underscores broader efforts within China’s tech sector to build a more self-sufficient AI ecosystem amid ongoing geopolitical and technological decoupling.
V4 succeeds the company’s V3 model released in late 2024, but it follows the breakout success of DeepSeek’s earlier R1 reasoning model, which drew global attention in early 2025 for its cost efficiency and competitive performance. That release was widely seen as evidence of China’s rapid progress in closing the AI gap with the United States.
According to DeepSeek, its top-tier “V4 Pro Max” model demonstrates strong performance across standard reasoning benchmarks, surpassing some earlier systems while trailing slightly behind the most advanced offerings from U.S. competitors. The company also claims its “pro” version approaches the capabilities of Anthropic’s most powerful models in agent-based tasks.
The lighter “flash” version, meanwhile, is designed for efficiency, delivering comparable performance in simpler tasks while maintaining strong reasoning capabilities. Both variants feature a context window of up to one million tokens, a substantial increase from the 128,000-token limit of V3, enabling the models to process and retain significantly larger volumes of information.
Industry observers say the new release positions DeepSeek as a serious contender in the global AI race. Analysts note that while V4 represents a solid advancement, it may not replicate the disruptive impact of the earlier R1 launch.
Huawei, in a separate statement, emphasized that its Ascend chips are fully compatible with DeepSeek’s latest models, highlighting the growing viability of a China-centric computing stack independent of Nvidia’s dominance.
Adoption of DeepSeek’s technology has already been expanding, particularly in developing markets where Huawei devices are widely used, according to earlier research. The company’s strategy of offering free access to its chatbot and maintaining an open-source approach has also helped attract developers seeking greater flexibility.
As competition between Chinese and U.S. AI firms intensifies, DeepSeek’s V4 debut signals both technological progress and a broader shift toward parallel innovation ecosystems, with each side seeking to define the future of artificial intelligence on its own terms.