DeepSeek Unveils New AI Model Powered by Huawei, Advancing China's Tech Independence
Beijing, Friday, 24 April 2026.
DeepSeek’s highly anticipated V4 AI model now runs on domestic Huawei chips. This breakthrough advances China’s tech self-sufficiency, potentially bypassing U.S. export controls and reshaping the global semiconductor landscape.
A Leap in Domestic Hardware Integration
Breaking months of silence, the Hangzhou-based artificial intelligence laboratory DeepSeek officially launched the preview of its next-generation foundational model, V4, between April 23 and April 24, 2026 [1][2][3][4]. The open-source model arrives in two distinct configurations: the massive V4-Pro, which boasts 1.6 trillion parameters, and the more lightweight V4-Flash, featuring 284 billion parameters [3][6]. This release directly answers a mandate from Beijing, reportedly issued in January 2026, which pushed domestic technology companies to adopt local semiconductor alternatives [1].
Architectural Innovations and Benchmark Triumphs
To achieve these results without relying exclusively on top-tier American hardware, DeepSeek implemented an architecture known as “Sparse Attention” [4][6]. This technique allows the neural network to focus selectively on the most relevant data rather than processing all information simultaneously, drastically reducing computational and memory requirements [4][6]. Consequently, both the Pro and Flash versions feature a massive context window of 1 million tokens, representing a staggering increase of 681.25% over the 128,000-token limit of DeepSeek’s previous flagship model [3][6].
Market Shockwaves and Intensifying Competition
The financial markets reacted immediately to the V4 preview on April 23, 2026, reflecting a clear divergence between domestic hardware manufacturers and competing software developers [1]. Shares of Chinese semiconductor foundries surged in Hong Kong trading, with Hua Hong Semiconductor jumping 15%, while Semiconductor Manufacturing International Corporation (SMIC) saw gains reported between 9% and 10% [1][2][alert! ‘Sources differ slightly on SMIC exact percentage gain, reporting either 9% or 10%’]. Conversely, domestic AI rivals faced sharp sell-offs; MiniMax and Knowledge Atlas Technology fell approximately 8%, and Manycore Tech plunged 9% [1][2].
The Future of the Open-Source Ecosystem
DeepSeek’s rapid iteration highlights a broader acceleration in global artificial intelligence development [1]. The company’s old application programming interface (API) models, deepseek-chat and deepseek-reasoner, are slated to be phased out by July 24, 2026, as developers are transitioned to the V4 ecosystem [6]. To support this rapid expansion, DeepSeek’s parent entity, High-Flyer Capital Management, is currently seeking to raise additional capital at a valuation exceeding $20 billion [2].