During the 2026 Spring Festival, while the much-discussed DeepSeek V4 large model was not officially released, DeepSeek quietly rolled out a version update on February 11, packed with numerous highlights. This update expanded the context length from 128K to 1M (one million Tokens), a nearly tenfold increase, enabling it to process approximately 300,000 lines of code or entire books, such as the complete collection of *The Three-Body Problem*, in one go. The new version's knowledge base has been updated to May 2025, with significant improvements in several core capabilities. It now supports a maximum context length of 1M Tokens, a rare feat among current models capable of handling contexts at the million-token level. Additionally, DeepSeek disclosed two innovative architectures: mHC (Manifold-Constrained Hyper-Connection) and Engram (Conditional Memory Module), which optimize information flow in deep Transformers and reduce long-context inference costs, respectively. Although the new version currently does not support visual input or multimodal recognition capabilities, its performance in programming and complex task processing already aligns with mainstream closed-source models. DeepSeek V4 is expected to be officially released in mid-February, featuring robust code generation capabilities and ultra-long context processing, poised to drive a new wave of technological iteration for domestic large models.