Chinese AI startup DeepSeek is preparing to launch its next flagship model, DeepSeek V4, expected in mid-February 2026 around Lunar New Year. Internal testing reportedly shows V4 outperforming Claude 3.5 Sonnet and GPT-4o on coding benchmarks, though independent verification is pending.
V4 introduces context windows exceeding 1 million tokens, enabling full codebase analysis in a single pass. The model can trace dependencies across multiple files and maintain consistency during large-scale refactoring. Engram conditional memory technology, published January 13, enables efficient retrieval from massive contexts.
DeepSeek's January 1 paper introduced mHC (Manifold-Constrained Hyper-Connections), improving scalability while reducing computational demands. Recent GitHub updates revealed "MODEL1" appearing 28 times across 114 files, hinting at the new architecture.
