DeepSeek released and open-sourced its V4 preview model. It features a one-million-token context and leads open-source models in Agent capabilities, reasoning, world knowledge, math, STEM, and competitive coding — rivaling top closed-source models. V4-Pro boosts AgenticCoding, V4-Flash is faster and cheaper, and the new DSA sparse attention mechanism sharply cuts compute and memory use. DeepSeek vows to stick to long-termism on its path to AGI.

By Wang Xinyuan, Liu Shujun(intern)/China Youth Daily

Source: The Paper

打开网易新闻 查看精彩图片

来源：中国青年报客户端