DeepSeek Quietly Updates Website with New Model: 1M Token Context and May 2025 Knowledge Cutoff
DeepSeek has made a subtle update to its official website, introducing what appears to be a new model. While the exact version name isn't explicitly stated, testing reveals it is highly unlikely to be the previously rumored "DeepSeek-V3.2," but rather an entirely new model with distinct characteristics.
Verified by DataLearnerAI, this new model demonstrates clear behavioral differences compared to earlier versions.
Context Window and Knowledge Cutoff Evolution
The most striking discovery comes from the model's own responses: it supports up to 1 million tokens for input and claims a knowledge cutoff date of May 2025. Previously, DeepSeek had ceased officially disclosing knowledge cutoff dates since the V3.1 release, with the DeepSeek-V3-0324 version citing a cutoff of July 2024.
The emergence of such a recent cutoff strongly hints at a major internal overhaul or a shift toward a next-generation model, potentially DeepSeek-V4. This leap in context window capacity and data recency provides significant advantages for AI agent applications and complex long-document analysis. Although an official announcement from DeepSeek is still pending, users can already access and benefit from these enhancements via the official site.
Related Articles
Loading...