DeepSeek has released its newest artificial intelligence (AI) model, DeepSeek-V4, promising drastically lower computing and memory costs for users and developers.
DeepSeek-V4 supports an ultra-long context window of one million words, which the Hangzhou-based company described as world-leading across both domestic and open-source fields.
The model is available in two versions: DeepSeek-V4-Pro, with 1.6 trillion parameters, and DeepSeek-V4-Flash, carrying a smaller 284 billion parameters for efficiency.
DeepSeek-V4-Flash is positioned as the more economical choice, offering faster and lighter performance while the Pro version targets advanced reasoning and world knowledge tasks.
A preview version of the open-source model is now publicly available, the company confirmed via posts on WeChat and the social media platform X on Friday.
The model has been optimised for AI agent tools including Claude Code, OpenClaw, OpenCode, and CodeBuddy, according to the official statement released by DeepSeek.
Performance and industry impact
In world knowledge benchmarks, DeepSeek-V4-Pro leads all open-source rivals and trails only the Gemini-Pro-3.1 of Google among the top closed-source models available today.
Tech research founder Zhang Yi told AFP that V4 marks a genuine inflection point, addressing slower performance and higher costs long associated with extended context lengths.
Zhang added that ultra-long context support, once limited to high-end research labs, is now expected to enter mainstream commercial applications for everyday business use.
Background and rivalry
DeepSeek first stunned the global technology industry in January 2024, when its R1 reasoning model matched leading US systems at a fraction of their reported development cost.
That release triggered a sharp sell-off in AI-related stocks worldwide and was widely described as a Sputnik moment for the artificial intelligence industry internationally.
DeepSeek’s systems are open source, making their internal workings publicly accessible, which contrasts sharply with the proprietary models sold by OpenAI and other Western rivals.
The company’s tools have since been adopted by Chinese municipalities, healthcare institutions, financial sector firms, and other businesses across the country’s domestic market.
US–China tensions
The White House on Thursday accused Chinese entities of running industrial-scale campaigns to steal American artificial intelligence technology through a practice known as distillation.
Michael Kratsios, science and technology chief advisor to President Donald Trump, said on X that foreign entities, primarily in China, were systematically targeting US AI systems.
Distillation is a recognised practice within AI development, commonly used to create smaller, cheaper model versions, but the White House framed China’s use as theft.
The accusations came ahead of an expected summit between President Trump and China’s President Xi Jinping, scheduled to take place in Beijing next month.
Wider industry context
The Friday announcement from DeepSeek coincided with Meta confirming plans to cut approximately one tenth of its global workforce as part of a broader productivity and AI push.
Reports also indicated that Microsoft was examining workforce reductions, as major technology companies continue to shift resources and investment heavily toward artificial intelligence development.
