DeepSeek silently expands context window to 1M tokens and updates knowledge cutoff to May 2025

DeepSeek pushed an unannounced update to its existing models on February 11, expanding the context window from 128K to 1 million tokens and moving the knowledge cutoff forward to May 2025. No changelog or announcement was published. The change was surfaced by developers comparing API behavior before and after. Community interpretation: DeepSeek is staging infrastructure for the V4 release, which has been anticipated since late 2025.

The silent rollout is characteristic of DeepSeek's operating style — ship quietly, let results speak. The 1M context expansion puts DeepSeek's existing models in the same tier as Gemini and Opus 4.6 on raw context capability. More importantly, it signals V4 is closer than the prolonged silence has suggested. DeepSeek V3's surprise release in late 2024 rattled markets; V4 with a larger MoE architecture and the reported Engram memory system could do it again.