HANGZHOU, April 24, 2026 — China’s DeepSeek is back.
More than a year after its R1 model blindsided Silicon Valley and sent shockwaves through global stock markets, the Hangzhou-based AI startup has unveiled its most ambitious release yet. DeepSeek launched preview versions of its V4 model Friday, promising drastically reduced costs and capabilities that once again challenge the dominance of US rivals.
The new release comes in two flavours. DeepSeek-V4-Pro packs a staggering 1.6 trillion parameters, while the leaner DeepSeek-V4-Flash comes in at 284 billion — both sporting a 1 million-token context window, allowing entire codebases or lengthy documents to be processed in a single prompt.
DeepSeek also highlighted a breakthrough it calls Hybrid Attention Architecture, which the company claims significantly improves the model’s ability to retain context across long conversations — a key advancement for enterprise and developer applications.
On the performance front, DeepSeek hasn’t held back on ambition. The company claims V4-Pro Max delivers superior performance on standard reasoning benchmarks compared to OpenAI’s GPT-5.2 and Google’s Gemini 3.0-Pro, though it falls marginally short of the very latest closed-source flagships from both rivals.
Perhaps most strikingly, V4 was trained using domestic chips from Huawei and Cambricon — a significant departure from R1, which relied on Nvidia hardware. Huawei confirmed its “Ascend 950” Supernode technology underpins the new models, a development analysts say could carry even broader implications than the original DeepSeek shock. “It allows AI systems to be built and deployed without relying solely on Nvidia,” said Wei Sun, principal analyst at Counterpoint Research, suggesting V4 could accelerate global AI adoption far beyond what R1 achieved.
Yet analysts are urging restraint. “R1 shocked US markets because no one expected a Chinese model to compete at that level. V4 is simply a follow-through on that same trend, and trends don’t make headlines the way shocks do,” said Ivan Su, senior equity analyst at Morningstar.
Shares of several Chinese AI competitors fell in Hong Kong trading on Friday, with MiniMax and Zhipu each dropping around 8%, and Hangzhou-based Manycore Tech plunging 9% — a sign that domestic AI competition in China has intensified sharply since R1’s breakthrough moment.
The release lands amid escalating geopolitical tensions. The White House this week accused Chinese entities of running “industrial-scale distillation campaigns” to steal American AI technology — allegations DeepSeek has not formally addressed. The announcement also came just days after reports emerged that Tencent and Alibaba were in talks to invest in DeepSeek at a valuation exceeding $20 billion.
For now, DeepSeek’s V4 is text-only, with the company confirming it is actively working on multimodal capabilities that would allow the model to process images and video — a signal that the startup’s ambitions are far from finished.
The world is watching — again.
Discover more from News
Subscribe to get the latest posts sent to your email.