DeepSeek launched two versions of its V4 model on Friday: the V4-Pro with 1.6 trillion parameters and the smaller V4-Flash with 284 billion parameters. Both models feature a 1 million token context window, achieved with what the company describes as "world-leading" cost efficiency.

The performance benchmarks tell a striking story. According to DeepSeek's announcement, V4-Pro beats all rival open models for mathematics and coding, trailing only Google's closed Gemini 3.1-Pro for world knowledge. The company claims its performance falls "marginally short" of OpenAI's GPT-5.4 and Gemini 3.1-Pro, "suggesting a developmental trajectory that trails state-of-the-art frontier models by approximately 3 to 6 months."

"Based on the benchmark results, it does appear DeepSeek V4 is going to be very competitive against its U.S. rivals," Lian Jye Su, chief analyst at technology research group Omdia, told Greenwich Time.

Hardware Independence Strategy

The V4 launch demonstrates a crucial shift in DeepSeek's infrastructure strategy. While previous models relied heavily on Nvidia chips, the company worked closely with Huawei to optimize V4 for China's domestic Ascend AI chip line. Huawei announced "full support" for V4 models across its range of Ascend processors and supernode systems.

This hardware collaboration addresses one of China's most pressing AI challenges — reducing dependence on US semiconductor technology amid tightening export controls. The partnership allows DeepSeek to maintain cutting-edge performance while operating within China's domestic supply chain.

The V4-Flash model maintains pricing identical to DeepSeek's V2 model from June 2024, making it one of the cheapest cutting-edge models available globally. This aggressive pricing strategy could pressure US competitors who rely on higher margins to fund their massive infrastructure investments.

Distillation Controversy

DeepSeek's rapid progress has drawn scrutiny from US AI leaders. In February, Anthropic accused DeepSeek and two other Chinese AI laboratories of running "industrial-scale campaigns" to "illicitly extract Claude's capabilities to improve their own models" through a technique called distillation — training weaker models on outputs from stronger ones.

OpenAI made similar allegations in a letter to US lawmakers, suggesting Chinese companies were reverse-engineering American AI capabilities. DeepSeek has not publicly responded to these specific accusations.

The distillation debate highlights a fundamental challenge in AI competition: when model outputs are publicly available through APIs, preventing knowledge transfer becomes nearly impossible to enforce through traditional export controls.

Global Implications

Marina Zhang, an associate professor at the University of Technology Sydney, described the V4 rollout as a "pivotal milestone for China's AI industry," especially as global competition intensifies in the pursuit of technological self-reliance.

The timing appears deliberate. DeepSeek's previous R1 reasoning model, released in January 2025, "rocked global tech markets" according to CNBC, raising questions about the scale of AI infrastructure spending when competitive open-source alternatives emerge.

The V4 release comes more than a year after R1's viral success, suggesting DeepSeek has used that time to build sustainable development capabilities rather than chase quick market attention. The company's methodical approach — developing both efficient training methods and hardware partnerships — positions it as a long-term competitor rather than a one-time disruptor.

For US policymakers betting billions on export controls as an AI containment strategy, DeepSeek's V4 represents a stark reality check. When frontier-level AI capabilities can be packaged as downloadable open-source software, controlling the hardware becomes significantly less effective at controlling the technology itself.