Chinese AI lab DeepSeek has released V4, its first major model since the R1 reasoning breakthrough in January 2025. The trillion-parameter multimodal model uses a Mixture-of-Experts architecture with roughly 32 billion parameters active per token, features a 1-million-token context window, and can natively process and generate text, images, and video — all under an Apache 2.0-style open-source license.

The release carries a sharp geopolitical edge. DeepSeek optimized V4 for Huawei Ascend and Cambricon chips and deliberately withheld early access from Nvidia and AMD, breaking with standard industry practice. Priced at approximately $0.14 per million input tokens — roughly one-twentieth the cost of GPT-5 — V4 positions itself as a serious open-source alternative to Western frontier models.

The launch was timed to coincide with China's annual Two Sessions parliamentary meetings, reinforcing DeepSeek's role as a national AI champion. Since R1, competitors like Alibaba had captured demand for affordable open-source AI, but V4 reasserts DeepSeek's position at the frontier of cost-efficient, high-performance models.