Chinese artificial intelligence startup DeepSeek has unveiled its latest models, DeepSeek-V4-Pro and DeepSeek-V4-Flash, marking the Hangzhou-based company's most significant release since it stunned the global technology sector more than a year ago. The new models feature an ultra-long context window of one million words — a measure of how much input a model can process at once — along with what the company describes as "drastically reduced" computing and memory costs. Preview versions of both open-source models are now publicly available.

DeepSeek-V4-Pro, which carries 1.6 trillion parameters — the numerical values that shape a model's decision-making — leads all rival open-source models in mathematics and coding benchmarks, according to DeepSeek. It trails only Google's closed-source Gemini 3.1-Pro in world knowledge tasks and falls "marginally short" of OpenAI's GPT-5.4, with the company acknowledging its trajectory lags state-of-the-art frontier models by roughly three to six months. The smaller V4-Flash model, with 284 billion parameters, offers comparable reasoning at faster speeds and lower cost, making it suited to commercial deployment. Both models have been optimised for popular AI agent tools including Claude Code and CodeBuddy. Tech analyst Zhang Yi, founder of research firm iiMedia, described the release as a genuine "inflection point" for the industry, predicting that ultra-long context support will move beyond specialist research settings and into mainstream applications.

The release arrives against a backdrop of intensifying US-China rivalry over artificial intelligence. The White House on Thursday accused Chinese entities of conducting "industrial-scale" campaigns to extract capabilities from American AI models through a process known as distillation — training smaller models on the outputs of larger ones. Michael Kratsios, the White House science and technology adviser, said tens of thousands of proxy accounts and jailbreaking techniques were being used to systematically exploit American expertise. The Chinese Embassy in Washington rejected the allegations as "baseless," stating that Beijing takes intellectual property protection seriously. The accusations come weeks before a scheduled summit between President Donald Trump and Chinese President Xi Jinping in Beijing.

DeepSeek first made global headlines in January last year when its R1 reasoning model matched the performance of ChatGPT and Google Gemini while reportedly costing less than six million dollars to develop — a fraction of the multibillion-dollar budgets typical in Silicon Valley. The release, described by prominent venture capitalist Marc Andreessen as "AI's Sputnik moment," triggered a sharp sell-off in AI-related stocks and forced a reckoning among Western tech firms. Some analysts questioned whether DeepSeek had access to greater resources than it disclosed. Despite its technical achievements, DeepSeek drew criticism over data privacy and its refusal to engage with politically sensitive topics, prompting bans or restrictions in the United States, Australia, Taiwan, South Korea and several European countries. The new release lands as both Meta and Microsoft are reported to be trimming their workforces, even as they pour investment into artificial intelligence — underscoring how DeepSeek's cost-efficient approach continues to reshape industry assumptions about what AI development requires.