Post

On April 24, 2026, DeepSeek officially disrupted the frontier AI market with the release of the V4 series. This launch introduces two primary models: DeepSeek-V4-Pro, a 1.6 trillion parameter flagship, and DeepSeek-V4-Flash, a high-speed efficiency model. Both models are available as open weights under the MIT license, reinforcing DeepSeek's commitment to accessible high-tier intelligence. Technical Architecture and Innovation The V4 series introduces a significant architectural evolution from the previous V3.2 generation. Both models utilize a Mixture-of-Experts (MoE) structure, but the core efficiency gains come from two specific innovations: 1. DeepSeek Sparse Attention (DSA): This novel attention mechanism allows the models...