DeepSeek, which rattled global markets last year with its low-cost R1 reasoning model, has officially unveiled its next major leap, DeepSeek-V4.
The release marks a strategic pivot for the Chinese artificial intelligence (AI) startup: it is moving away from a reliance on U.S. silicon toward a deeper integration with domestic hardware.
The V4 model, currently in a preview release under an MIT open-source license, has been optimized to run on Huawei’s Ascend AI processors.
The collaboration represents a significant milestone for China’s AI sovereignty, as U.S. export controls continue to restrict the flow of high-end NVIDIA Corp. chips into the country.
Huawei confirmed its Ascend-powered computing clusters are fully capable of supporting the new model. While the exact ratio of NVIDIA-to-Huawei hardware used during the training phase remains undisclosed, analysts suggest that V4’s native compatibility with local chips could accelerate China’s ability to scale AI without Western infrastructure.
DeepSeek is offering the V4 in two distinct versions to balance performance and cost. DeepSeek-V4-Pro is a “mixture-of-experts” powerhouse featuring 1.6 trillion total parameters. According to the company, it trails only Google’s Gemini-3.1-Pro in world-knowledge benchmarks.
DeepSeek-V4-Flash, meanwhile, is a leaner version designed for speed and efficiency, sporting 284 billion parameters.
Both models feature a massive one-million-token context window, allowing them to process vast amounts of data in a single prompt. DeepSeek is positioning V4 as a specialized tool for agent-based tasks, designed to integrate seamlessly with coding tools like Anthropic’s Claude Code.
The debut sent immediate ripples through the financial sector.
In Hong Kong, domestic chipmakers SMIC and Hua Hong Semiconductor saw shares surge by 10% and 15%, respectively, as investors bet on a future powered by local hardware. Conversely, domestic AI rivals like Zhipu AI and MiniMax saw shares dip nearly 9%, reflecting a crowded and increasingly aggressive Chinese AI market.
The geopolitical weight of this release was perhaps best summarized by NVIDIA CEO Jensen Huang. Speaking recently on the Dwarkesh Podcast, Huang warned that the day a top-tier model like DeepSeek launches natively on Huawei hardware first would be a “horrible outcome” for U.S. tech leadership.
With V4 now live via API, the gap between American and Chinese AI performance appears to be closing, shifting the focus from who has the most money to who can do the most with the hardware they have.
