DeepSeek returned Friday with another model designed to make American AI executives uncomfortable, releasing DeepSeek-V4 with what the Chinese startup described as “drastically reduced” compute and memory costs — arriving more than a year after its original R1 model detonated assumptions about US dominance in artificial intelligence and wiped billions from the market valuations of companies that had built their investor cases on the premise that cutting-edge AI required cutting-edge spending.
The Hangzhou-based company said V4 features an ultra-long context window of one million words — meaning the model can absorb and process vastly more input at once than most current systems — while achieving what it described as world-leading performance across agent capabilities, world knowledge and reasoning. Two versions are available: DeepSeek-V4-Pro, carrying 1.6 trillion parameters, and DeepSeek-V4-Flash, a leaner 284-billion-parameter version positioned as the more economical option. A preview of the open-source model is now accessible to developers.
Context length — the amount of information a model can hold in view while completing a task — has been one of the persistent constraints of practical AI deployment. Systems that lose the thread of long documents, extended conversations or complex multi-step tasks require workarounds that add cost and reduce reliability. V4’s claimed million-word context, if it performs as described, would push long-text processing out of high-end research environments and into mainstream commercial use.
“This addresses the long-standing issues of slower performance and higher costs associated with long context lengths, marking a genuine inflection point for the industry,” said Zhang Yi, founder of tech research firm iiMedia. “For end users, this will bring widespread, accessible benefits.”
The announcement landed on a day when Meta confirmed it was cutting approximately 8,000 jobs — ten percent of its workforce — to fund escalating AI infrastructure costs, and Microsoft was reported to be offering voluntary buyouts to thousands of its American employees for similar reasons. The juxtaposition was pointed: Western technology giants spending more and employing fewer to stay competitive in an AI race, while a Chinese startup released a model claiming to achieve comparable or superior performance at lower cost.
DeepSeek-V4-Pro has been benchmarked against the field in world knowledge tests and found to significantly lead other open-source models, falling only slightly behind Google’s Gemini-Pro-3.1 among closed-source competitors. The model has been optimised for popular AI agent products including Claude Code, OpenCode and CodeBuddy — a deliberate integration play designed to embed V4 into the development workflows that engineers already use.
When DeepSeek’s R1 model appeared in January last year, it triggered what analysts and executives called a Sputnik moment — a sudden, disorienting recognition that the gap between American AI capability and Chinese AI capability was smaller than the industry had assumed, achieved at a fraction of the cost that US firms had spent. The original DeepSeek shock sent AI-related shares into a sharp sell-off as investors recalculated the return on the hundreds of billions being poured into infrastructure and talent by OpenAI, Google, Microsoft and Meta.
Read also: Meta Plans Job Cuts As Microsoft Offers Exit Packages
Friday’s V4 release arrives in a geopolitical context that has grown considerably more charged since then. The White House accused Chinese entities this week of running “industrial-scale distillation campaigns to steal American AI,” with Trump’s science and technology chief advisor Michael Kratsios posting the accusation on X. Distillation — the practice of training smaller, cheaper models using outputs from larger ones — is standard within AI development, but American officials have framed Chinese use of the technique as technology theft rather than engineering. The accusation comes ahead of an expected summit between Trump and Xi Jinping next month, ensuring that AI will be on the diplomatic agenda alongside trade and the Iran war.
DeepSeek’s decision to make its systems open-source has driven broad adoption inside China, with municipalities, healthcare institutions, financial sector firms and other businesses integrating its tools — partly because open-source access removes the licensing costs and dependency relationships that come with proprietary Western alternatives. That adoption base gives DeepSeek a commercial flywheel that feeds improvement back into subsequent model versions.
The questions that accompanied R1’s arrival have not been resolved. DeepSeek’s chatbot has consistently declined to engage with politically sensitive topics — the 1989 Tiananmen Square crackdown among them — raising concerns about censorship baked into systems that are simultaneously being positioned as open and globally accessible. Data privacy questions about a Chinese-developed system used at scale outside China have not been definitively answered.
V4 is faster, cheaper and longer-seeing than its predecessor. The Sputnik moment, it appears, was not a single event but a recurring condition.