DeepSeek V4 was widely expected to launch in mid-February 2026, but as of early March there has been no official announcement. The silence is unusual for a lab that has been aggressive about releases, and it has fuelled speculation about delays, regulatory complications, or a strategic decision to hold back until the model is significantly ahead of the competition.
What has appeared, quietly, is a context window expansion. DeepSeek's existing models now support up to 1M tokens with no formal changelog or announcement. Developers noticed the change when API calls that previously failed at around 128K tokens started succeeding with much larger payloads. The stealth update suggests active infrastructure work is happening behind the scenes, even if V4 itself remains unreleased.
Adding to the intrigue, a mystery model appeared on OpenRouter that several developers believe is an early V4 variant. The model shows improved reasoning, better code generation, and handles complex multi-step tasks more reliably than the current V3 release. DeepSeek has not confirmed or denied the connection, leaving the developer community to speculate based on benchmark results and behavioural analysis.
