DeepSeek Launches V4 AI Model Optimized for Huawei Chips

The Chinese AI startup DeepSeek released preview versions of its V4 open-source model on April 24-25, 2026. The rollout includes a 1.6-trillion parameter V4-Pro for high-level reasoning in STEM and coding, and a 284-billion parameter V4-Flash version. Both models feature a 1 million token context window and agentic capabilities for autonomous code writing. To circumvent U.S. export controls, DeepSeek collaborated with Huawei and Cambricon Technologies to rewrite underlying code for domestic hardware, specifically utilizing Huawei Ascend 950 chips and Supernode technology.

DeepSeek claims V4-Pro rivals top closed-source systems, trailing only Google's Gemini-Pro-3.1 in world knowledge. To attract users, senior researcher Victor Chen announced aggressive pricing, including a permanent 90% discount for cache hits and a 75% base promotion, making some outputs up to 35 times cheaper than OpenAI's GPT-5.5. However, analysts from Omdia and Ankura China Advisors noted a muted market reaction compared to previous releases, suggesting such efficiencies are now priced into valuations.

The launch triggered significant geopolitical friction. The White House, OpenAI, and Anthropic accused DeepSeek of using "distillation"—illicitly extracting capabilities from U.S. models using thousands of fake accounts. U.S. officials also alleged that V4 was trained on banned Nvidia Blackwell chips. The Government of China dismissed these claims as unjustified suppression. Nvidia CEO Jensen Huang described the pivot to Huawei hardware as a "horrible outcome" for the U.S. due to the potential loss of the Chinese developer ecosystem.