DeepSeek, a prominent Chinese artificial intelligence startup, has launched preview versions of its latest major update, DeepSeek V4, amidst an intensifying AI rivalry between China and the U.S. This highly anticipated release builds upon its predecessor, V3.2, and follows the impactful R1 reasoning model that garnered significant attention upon its release in January 2025 for its cost-effectiveness and symbolic representation of China's technological advancements.

DeepSeek V4 introduces two primary open-source models: 'Pro' and 'Flash' versions, both leveraging a mixture-of-experts architecture. A key enhancement is their expanded context window of 1 million tokens, a substantial increase from V3's 128,000 tokens, enabling them to process large codebases or documents efficiently. The Pro model boasts an impressive 1.6 trillion parameters (with 49 billion active), making it the largest open-weight model available, surpassing rivals like Moonshot AI’s Kimi K 2.6 (1.1 trillion) and MiniMax’s M1 (456 billion), and more than doubling DeepSeek V3.2’s 671 billion parameters. The smaller V4 Flash version features 284 billion parameters (13 billion active).

The company claims significant improvements in knowledge, reasoning, and 'agentic' capabilities—the ability to perform complex tasks autonomously. DeepSeek states that the V4 models are more efficient and performant than DeepSeek V3.2 due to architectural enhancements, and have 'almost closed the gap' with leading open and closed models on reasoning benchmarks. Specifically, the 'V4 Pro Max' version demonstrates superior performance against OpenAI’s GPT-5.2 and Google’s Gemini 3.0-Pro on standard reasoning benchmarks, falling only marginally short of GPT-5.4 and Gemini 3.1-Pro. In terms of agentic capabilities, the V4 'Pro' version is said to outperform Claude’s Sonnet 4.5 and approach the level of Claude’s Opus 4.5, based on DeepSeek’s own evaluations. Both V4 models' performance in coding competition benchmarks is deemed 'comparable to GPT-5.4.'

Despite these advancements, DeepSeek acknowledges that its V4 models fall slightly behind frontier models like OpenAI’s GPT-5.4 and Google’s Gemini 3.1 Pro in knowledge tests, suggesting a developmental lag of approximately '3 to 6 months.' A notable limitation is that both V4 Flash and V4 Pro currently support text only, unlike many of their closed-source competitors that offer multi-modal capabilities encompassing audio, video, and image understanding and generation.

A significant advantage of DeepSeek V4 is its affordability. The smaller V4 Flash model is priced at $0.14 per million input tokens and $0.28 per million output tokens, significantly undercutting models like GPT-5.4 Nano, Gemini 3.1 Flash, GPT-5.4 Mini, and Claude Haiku 4.5. The larger V4 Pro model is also cost-effective at $0.145 per million input tokens and $3.48 per million output tokens, offering a more affordable alternative to Gemini 3.1 Pro, GPT-5.5, Claude Opus 4.7, and GPT-5.4. DeepSeek also positions its technology as 'open source,' providing developers access to modify and build on its core, a distinction from the proprietary nature of top models from Anthropic, Google, and OpenAI.

The launch of DeepSeek V4 comes amidst serious accusations. Anthropic and OpenAI have accused DeepSeek, along with other China-based AI laboratories, of 'industrial-scale campaigns' to 'illicitly extract capabilities' from their models through a technique called 'distillation.' This involves training a less capable model on the outputs of a stronger one. Michael Kratsios, chief science and technology adviser to U.S. President Donald Trump, echoed these concerns, accusing foreign tech companies, 'principally based in China,' of 'exploiting American expertise and innovation.' China’s embassy in Washington has vehemently denied these allegations, calling them 'unjustified suppression of Chinese companies by the U.S.'

Industry analysts have offered varied perspectives. Lian Jye Su, chief analyst at Omdia, believes DeepSeek V4 is 'going to be very competitive against its U.S. rivals.' Marina Zhang, an associate professor at the University of Technology Sydney, hails V4's rollout as a 'pivotal milestone for China’s AI industry,' especially in the global pursuit of self-reliance in critical technologies. However, Ivan Su, a senior equity analyst at Morningstar, views V4 as a 'competent' follow-up but not as groundbreaking as R1, citing intensified domestic competition. He emphasizes the need for independent evaluations before drawing final conclusions on its capabilities compared to U.S. models, despite DeepSeek’s own evaluations suggesting comparable performance on most fronts. A report from Microsoft in January indicated DeepSeek's growing adoption in many developing nations, underscoring its expanding global presence.