Skip to content
X · @dotey

DeepSeek releases and open-sources V4 preview; 1M context is standard across all services

DeepSeek 今天发布了全新的 V4 系列模型预览版,同步开源。最大的变化是把百万(1M)上下文直接变成了所有官方服务的标配,不分版本、不分价位。

DeepSeek released and open-sourced the V4 preview, making 1M context standard across all official services with no tier or price split. The post says V4-Pro and V4-Flash use token compression plus DSA sparse attention to cut compute and memory costs for 1M context; legacy APIs remain for 3 months and stop after July 24.

Why it matters: DeepSeek is a flagship Chinese model vendor, and this V4 preview is a substantive release with open source and 1M context made standard across official services. HKR-H/K/R all pass: the post includes mechanisms and a migration deadline, and the tier reset makes it a same-day P1.

Read the original ↗Export Markdown