DeepSeek-V4 arrives with near SotA intelligence at 1/6th the cost (venturebeat.com)

DeepSeek, the Chinese AI startup, has unveiled DeepSeek-V4, a significant advancement in AI models that boasts 1.6 trillion parameters and a Mixture-of-Experts architecture. Released under a commercially-friendly MIT license, this model approaches, and in some benchmarks even surpasses, the performance of leading proprietary systems while costing approximately one-sixth as much. This launch is being referred to as the "second DeepSeek moment", marking a pivotal shift in the economics of AI deployment by making advanced models much more accessible to businesses. The DeepSeek-V4 model achieves these impressive results through innovative technical advancements, including a native one-million-token context window and a Hybrid Attention Architecture that drastically reduces the memory footprint. It incorporates a unique Manifold-Constrained Hyper-Connections design to enhance stability and signal propagation during training, alongside a Mixture-of-Experts training approach that selectively activates only a fraction of its parameters. While benchmarking against competitors like GPT-5.5 and Claude Opus 4.7 indicates that DeepSeek-V4 still trails in some areas, its cost-effectiveness—especially with options like the Flash variant—compels enterprises to reevaluate the viability of automating tasks that previously seemed too costly, potentially reshaping the competitive landscape in AI technologies.