Three reasons why DeepSeek’s new model V4 matters

DeepSeek released a preview of V4, its most significant model since R1. The release has three key implications: it offers frontier-level performance at a fraction of the cost of OpenAI and Anthropic models, with V4-Pro at $1.74/million input tokens; it introduces a new attention mechanism that dramatically reduces memory and compute costs for 1-million-token context windows (using only 27% of the compute of V3.2); and it marks DeepSeek's first model optimized for Chinese domestic chips like Huawei's Ascend, signaling China's push toward AI hardware independence from Nvidia. V4 comes in two variants — V4-Pro for complex coding and agent tasks, and V4-Flash for speed and cost efficiency. While DeepSeek still appears to rely partly on Nvidia for training, the inference workload is being shifted to Chinese chips, with prices expected to drop further once Huawei's Ascend 950 supernodes ship at scale.