DeepSeek Quietly Tests New Model: 1M Token Context, Is V4 On The Way?

Recently, DeepSeek's App and web versions started gray testing a new model, with the context window length increased from V3.1's 128K tokens to 1M (one million) tokens, and the knowledge cutoff date updated to May 2025. However, it does not support visual input or multi-modal recognition yet.\nEarlier, a mysterious model codenamed "Model1" was found in the update of the GitHub repository FlashMLA. Its architecture is significantly different from V3.2, indicating a major restructuring, which echoes the news that DeepSeek plans to release V4.

The current current test test version is faster faster than the V3 series, but details such as specific parameters and whether it integrates new architecture components have not been disclosed, possibly a prelude to V4.