Three reasons why DeepSeek’s new model matters (www.technologyreview.com)

On April 24, DeepSeek, a Chinese AI company, unveiled V4, its new flagship open-source model, which can process significantly longer prompts compared to its predecessor, R1. V4 offers two versions: V4-Pro, optimized for complex tasks, and V4-Flash, designed for speed and cost-efficiency. Notably, V4-Pro's pricing is significantly lower than similar models from competitors, making it an attractive option for developers seeking affordable access to cutting-edge AI technology. Performance benchmarks indicate that V4-Pro rivals leading closed-source models like OpenAI’s GPT-5.4 and Google’s Gemini-3.1 while exceeding many open-source alternatives, particularly in coding and STEM problem-solving capabilities. V4’s advancements are not limited to performance; it introduces a new approach to memory efficiency by utilizing a longer context window of up to 1 million tokens and optimizing the attention mechanism to reduce computational costs dramatically. Additionally, this release marks DeepSeek's first model optimized for domestic Chinese chips, such as those from Huawei, reflecting a strategic shift towards reducing reliance on Nvidia amid increasing government pressures for self-sufficiency. While V4 may not have the immediate impact of R1, its innovative architecture and economic pricing could profoundly influence the accessibility of AI technologies in the rapidly evolving landscape of the AI/ML community.