Technology & Science
DeepSeek Debuts V4 LLM Preview Optimized for Huawei Ascend Chips
On 24 April 2026 DeepSeek released preview versions of its V4 language model—its first flagship tuned for Huawei’s Ascend 950 hardware—signaling a pivot away from Nvidia dependence while slashing usage prices.
Focusing Facts
- V4-Pro packs 1.6 trillion parameters (49 billion active) and supports a 1-million-token context window, making it the biggest open-weight model yet disclosed.
- Pricing: V4-Flash costs $0.28 and V4-Pro $3.48 per million output tokens versus OpenAI’s $30 and Anthropic’s $25.
- Hong Kong–listed SMIC and Huahong Semiconductor jumped 10–15% while rival model makers Zhipu AI and MiniMax fell 9% after the announcement.
Context
China’s quest for tech self-reliance echoes Japan’s 1982–1992 Fifth Generation Computer Project—an attempt to leapfrog U.S. computing that spurred American semiconductor policy even though the project ultimately stalled. Today’s U.S. export controls on Nvidia GPUs replay Cold-War era CoCom restrictions, pushing Chinese firms such as DeepSeek to co-develop with domestic chipmakers. The V4-Ascend pairing illustrates an accelerating bifurcation of the global AI stack: competing hardware–software ecosystems reminiscent of the IBM/Intel/Microsoft dominance fracture when ARM-based mobile computing rose after 2007. Whether V4 truly narrows the quality gap is secondary; what matters on a 100-year arc is the institutional learning—Chinese fabs, toolchains, and model developers are internalising capabilities that diminish leverage of any single foreign supplier. Like Sputnik (1957) jolted U.S. space investment despite limited immediate military shift, V4 on Huawei chips may similarly galvanise both sides into deeper, parallel innovation trajectories that entrench technological blocs for decades.
Perspectives
Chinese state-aligned media
e.g., China Daily, South China Morning Post — Portrays DeepSeek V4’s Huawei adaptation as a landmark step toward national tech self-reliance and a decisive break from dependence on U.S. chips. Driven to showcase China’s industrial success stories, the coverage downplays the model’s remaining performance gap with U.S. ‘frontier’ systems and glosses over export-control-related constraints.
International business & tech press
e.g., Reuters, Fortune, Yahoo News, BNN — Frames V4 as an impressive, cut-price open-source model that broadens Huawei’s ecosystem yet still lags top U.S. models by a few months on knowledge benchmarks. Market-oriented reporting can amplify DeepSeek’s own performance claims and benchmark charts to generate investor buzz, while giving limited scrutiny to questions of IP origin or national-security implications.
Security-focused Western commentary
e.g., The Japan Times quoting CFR analysts — Argues the preview shows DeepSeek has not closed the gap, offering reassurance that U.S. leadership in AI remains intact. By emphasizing American strategic dominance, this perspective may understate the significance of China’s price and hardware strides to reinforce a favorable policy narrative.
Like what you're reading?