The artificial intelligence race is accelerating. OpenAI launched GPT-5.5 this week, while the White House accused China of copying US AI systems at scale. Now, DeepSeek has introduced preview versions of its V4 model, targeting direct competition with top US platforms.

DeepSeek released the V4 Flash and V4 Pro series, highlighting gains in coding, reasoning, and agent-driven tasks. The models incorporate architectural upgrades and optimization improvements, with a clear focus on efficiency as systems grow more expensive to run.

A key feature is what DeepSeek calls Hybrid Attention Architecture. The method improves how models retain context across long conversations and reduces memory loss in extended interactions.

The system also supports a 1 million-token context window, allowing users to input entire codebases or long documents in a single prompt. This could reshape workflows in software development and enterprise analysis.

DeepSeek reported strong benchmark performance against systems from Anthropic, Google, and OpenAI. However, it acknowledged that V4 still trails the most advanced models by three to six months, while emphasizing cost and deployment flexibility.

Cost and chip strategy

DeepSeek continues to focus on efficiency as a competitive edge. Its trillion-parameter system uses a Mixture-of-Experts approach, activating only a fraction of parameters per task. This reduces inference costs compared to traditional models, which typically activate all parameters for each request.

The models are also designed to run on domestic hardware. DeepSeek expects costs to drop further once clusters powered by Huawei Technologies Co.’s Ascend 950 chips come online later this year. This shift could reduce reliance on US chipmakers and strengthen China’s AI infrastructure.

Markets reacted quickly. Shares of Semiconductor Manufacturing International Corp. and Hua Hong Semiconductor rose, while rival AI firms declined. Investors appear to be betting on increased demand for Chinese-made chips.

DeepSeek said service capacity for the V4 Pro series remains constrained due to limited computing resources. It is also in talks with Tencent Holdings Ltd. and Alibaba Group Holding Ltd. for its first funding round, signaling plans to expand infrastructure.

Pressure on US rivals

The V4 launch follows the earlier R1 model, which shook the AI market and prompted a reassessment of spending on frontier systems. DeepSeek claimed R1 delivered competitive performance at a fraction of the cost of leading US models.

That debate has since shifted again. US tech firms are projected to invest around $650 billion in 2026 on AI infrastructure and data centers, balancing performance gains with long-term costs.

DeepSeek said V4 builds on that approach by improving both scale and efficiency. It continues to position open-source models as alternatives to closed systems, appealing to developers and enterprises seeking more control.

Still, the release comes under scrutiny. US officials have accused DeepSeek of using restricted chips, while Anthropic has alleged misuse of its Claude system. DeepSeek has not disclosed training costs or hardware details for V4.

The launch highlights intensifying global competition. By focusing on lower costs, scalable performance, and hardware flexibility, DeepSeek is challenging how AI systems are built and who leads the next phase of development.