TL;DR:

• Chinese startup DeepSeek has launched a preview of its V4 model optimised for Huawei’s Ascend chip line, its first major release since the viral V3 launch of early 2025.
• The pro version trails only Google’s closed-source Gemini-Pro-3.1 on world-knowledge benchmarks, with rival Chinese AI stocks Zhipu AI and MiniMax falling 9% and 7% respectively on the news.
• The launch lands one day after the White House accused China of industrial-scale intellectual property theft from US AI labs, sharpening the geopolitical backdrop to the release.

DeepSeek, the Hangzhou-based AI startup that rattled global markets with its cost-efficient V3 model last year, has released a preview of the V4 explicitly tuned for Huawei’s Ascend AI accelerators. Huawei confirmed that its full Ascend supernode product line now supports the V4 series, and DeepSeek declined to disclose which chips were used for training.

A different silicon story to V3

V3 was trained on Nvidia hardware, and DeepSeek was subsequently accused by Washington of breaching US export controls to acquire cutting-edge Nvidia chips. The V4 announcement inverts the positioning: Huawei is front and centre as both collaborator and deployment target, while Nvidia is conspicuously absent. That matters commercially — the pro version beats every other open-source model on world-knowledge benchmarks, a result Chinese rivals clearly read as a threat given Zhipu AI’s 9% drop and MiniMax’s 7% slide on launch day. The preview also ships in a lower-cost flash variant, the same two-tier pattern DeepSeek used to undercut Western labs on V3.

DeepSeek is reportedly raising at a valuation above $20bn, with Alibaba and Tencent both said to be in talks for stakes. For UK enterprise teams evaluating open-weights models, V4’s benchmark profile — second only to a closed-source Google model — makes it harder to dismiss on capability grounds alone, even as the sovereignty and export-control story around Chinese labs grows noisier.

Geopolitics catching up with the release

The timing is pointed. White House Office of Science and Technology Policy director Michael Kratsios accused Chinese entities on Thursday of running “industrial-scale campaigns to distil US frontier AI systems” using tens of thousands of proxy accounts and jailbreaking techniques. DeepSeek was cited repeatedly in past US accusations of distilling proprietary OpenAI and Anthropic models. The Chinese Embassy in Washington rejected the claims, and Beijing’s foreign ministry urged the US to “abandon biases”.

Looking Forward

For UK businesses, V4 raises a practical sovereignty question: a genuinely frontier-adjacent open-weights model, trained on non-US compute, trailing only a closed Google model on benchmarks. Procurement conversations that focused on “Anthropic vs OpenAI” will now need to weigh DeepSeek’s capability, its chip-supply chain, and the tightening US policy posture. Trump’s summit with Xi Jinping next month — arriving shortly after the Kratsios memo — will shape how much of that calculation is controllable.