Friday’s previews of DeepSeek’s latest AI model update put the Chinese startup back at the center of the fast-moving race with U.S. rivals, with the company pitching improvements in reasoning and autonomous, task-executing behavior as well as changes aimed at reducing reliance on American chipmakers. The company said the update—its “V4” model—arrives with two preview versions, “pro” and “flash,” and that both carry a 1 million token context window.
The announcement comes after DeepSeek’s earlier models helped spark global market interest, including its specialized reasoning model R1, which was released in January 2025 and was framed by DeepSeek as more cost-effective than a comparable model from OpenAI. DeepSeek’s V4 is positioned as a successor to V3, which the company released in late 2024.
DeepSeek said the “V4 Pro Max” version delivers what it called “superior performance” on standard reasoning benchmarks relative to OpenAI’s GPT-5.2 and Google’s Gemini 3.0-Pro. In the same comparison, DeepSeek said V4 fell “marginally” short of GPT-5.4 and Gemini 3.1-Pro. The company’s marketing also leaned on what it called “agentic” capability, saying the “V4 pro” version could outperform Anthropic’s Claude 4.5’s Sonnet variant and approach Claude’s Opus 4.5 model based on DeepSeek’s own evaluation.
DeepSeek also presented “V4 flash” as a lower-cost option, saying it performs on a par with the “pro” version on simple agent tasks while maintaining reasoning capabilities that closely approach “pro.” DeepSeek’s product page positions both versions as open source in the sense that it enables developers to modify and build on the core technology, and the company provides a free-to-use web and mobile chatbot.
In addition to model comparisons, DeepSeek emphasized a key infrastructure shift: it said its V4 open-source models are supported in part by computer chips made by Huawei, reducing reliance on U.S. chipmakers such as Nvidia. Huawei said in a separate statement Friday that its Ascend chips and related technology are compatible with DeepSeek’s V4 models, describing the move as a demonstration of operating feasibility outside an Nvidia-dominated computing ecosystem.
Analysts said the V4 rollout is notable for what it signals about China’s push toward self-reliance, even as questions remain about whether the upgrade matches the scale of earlier breakthroughs. Lian Jye Su, chief analyst at the technology research and advisory group Omdia, said the benchmark results suggest DeepSeek V4 is “going to be very competitive against its U.S. rivals.” Marina Zhang, an associate professor at the University of Technology Sydney, called the rollout a pivotal milestone for China’s AI industry, particularly as global competition intensifies around critical technology independence.
Still, some skepticism surfaced around the size of the improvement, including how much of the progress can be confirmed by independent tests. Ivan Su, a senior equity analyst at Morningstar, said V4 is a “competent” follow-up but not as big a breakthrough as the launch of R1, and he said independent evaluations are needed before final conclusions can be drawn, noting that domestic competition has intensified since R1’s release.
The launch also lands amid renewed attention on how models are trained and improved, including disputes between major AI companies over alleged capability extraction. In February, Anthropic accused DeepSeek and two other China-based AI laboratories of “industrial-scale campaigns” to “illicitly extract Claude’s capabilities,” saying they used a technique called distillation that involves training a less capable model on outputs of a stronger one. OpenAI later made similar allegations in a letter to U.S. lawmakers.
This week, Michael Kratsios, chief science and technology adviser to U.S. President Donald Trump, accused foreign tech companies “principally based in China” of distilling leading U.S. AI systems and “exploiting American expertise and innovation.” China’s embassy in Washington responded by describing the allegations as “unjustified suppression of Chinese companies by the U.S.” As DeepSeek steps through the V4 preview phase, the debate over benchmarks and the underlying methods of model improvement is likely to remain central to how the update is assessed.
O’Brien reported from Providence, Rhode Island.