DeepSeek V4 Narrows Coding Gap With U.S. Leaders as China Doubles Down on Open Source

Chinese AI startup DeepSeek on Friday released preview versions of V4, its long-awaited flagship model, marking a significant advance in coding capabilities that narrows the performance gap with leading proprietary systems from Anthropic and OpenAI. The Hangzhou-based company unveiled both V4 Pro and V4 Flash variants as open-source releases, allowing developers worldwide to download, modify, and deploy the technology without licensing restrictions.

The V4 Pro model contains 1.6 trillion parameters—DeepSeek's largest to date—while the smaller V4 Flash version operates with 284 billion parameters. Both feature a one-million-token context window, a tenfold expansion over DeepSeek's previous flagship that enables processing of entire codebases or lengthy documents in a single prompt. According to tests conducted by Vals AI, a performance tracking firm, V4 significantly outperformed every other open-source system at generating computer code, though it still trails the top closed models from Anthropic's Claude and OpenAI's latest releases.

"Based on the benchmark results, it does appear DeepSeek V4 is going to be very competitive against its U.S. rivals," said Lian Jye Su, chief analyst at technology research group Omdia. Marina Zhang, an associate professor at the University of Technology Sydney, characterized the rollout as a "pivotal milestone for China's AI industry" as global competition intensifies around technological self-reliance.

The release arrived days after fellow Chinese startup Moonshot AI introduced its own open-source model, Kimi 2.6, underscoring the accelerating pace of domestic competition. DeepSeek now faces growing pressure from established players including Alibaba and ByteDance, both of which have released new models in recent months. The company's V4 launch comes more than a year after its R1 reasoning model disrupted global tech markets in January 2025 with claims of achieving competitive performance at a fraction of typical development costs.

(DeepSeek's V4 release explicitly highlights compatibility with Huawei's Ascend chip lineup and supernode systems, with the Shenzhen-based tech giant announcing "full support" for model inference shortly after Friday's unveiling. Domestic chipmaker Cambricon Technologies also moved quickly to confirm compatibility, signaling coordinated industrial policy backing.)

The strategic emphasis on open-source distribution stands in sharp contrast to the proprietary approach maintained by leading U.S. firms. OpenAI and Anthropic keep their most advanced models closed, restricting access and modification rights. DeepSeek's V3 model, released in late 2024, demonstrated that open-source systems could perform nearly as well as closed alternatives, triggering a wave of similar releases from Chinese competitors. By the end of 2025, Chinese open-source models had captured a significant share of global AI usage, according to industry tracking data.

The V4 rollout occurs against a backdrop of escalating technological rivalry between Washington and Beijing. This week the White House accused China of widespread technology theft in artificial intelligence, while OpenAI has alleged that DeepSeek used its proprietary models to train competing systems. The geopolitical dimension extends to chip access, with U.S. export controls limiting Chinese firms' ability to procure advanced semiconductors—a constraint DeepSeek has sought to overcome through architectural efficiency and domestic hardware partnerships.

DeepSeek's emphasis on "agentic" capabilities—the ability to autonomously perform complex workflows—reflects broader industry momentum toward AI systems that can execute multi-step tasks with minimal human oversight. The company claims its V4 Pro version outperforms Anthropic's Claude Sonnet 4.5 and approaches Claude Opus 4.5 levels in agent-based evaluations, based on internal testing. The Flash variant reportedly matches Pro performance on simpler agent tasks while maintaining close reasoning parity, offering a lower-cost deployment option.

The model's architecture incorporates what DeepSeek calls Hybrid Attention Architecture, designed to improve memory retention across extended conversational exchanges—a technical challenge that has constrained practical applications of large language models. The company also emphasized cost efficiency in achieving the expanded context window, positioning V4 as deployable on less expensive hardware than competing systems require, potentially lowering barriers for enterprise adoption in cost-sensitive markets.