Chinese AI lab DeepSeek is preparing to release V4, a one-trillion-parameter open-source model that will run on Huawei's Ascend chips instead of Nvidia hardware. Reuters confirmed on April 4 that this makes DeepSeek V4 the first frontier AI model built entirely on Chinese semiconductor infrastructure, directly challenging the assumptions underlying American chip export restrictions.
The model uses a Mixture-of-Experts architecture with approximately 37 billion active parameters per token and features a one-million-token context window powered by a novel Engram Conditional Memory system. Early benchmark leaks suggest SWE-bench scores above 80% and HumanEval at 90%, putting it competitive with leading Western models. Training reportedly cost just $5.2 million — a fraction of the $100 million-plus budgets typical at this scale.
DeepSeek plans to release V4 under an Apache 2.0 license with API pricing around $0.28 per million tokens, dramatically undercutting closed-source competitors. The combination of frontier-level performance, open weights, and independence from Western chips marks a significant shift in the global AI landscape.
