DeepSeek API Upgrades, V4 Version May be Incoming

Chinese AI company DeepSeek quietly expanded its API's context window from 128k Tokens to 1 million Tokens without releasing any official announcement on April 22, while also updating the knowledge cutoff date to May 2025. Several developers have confirmed the above changes on social media.

This upgrade brings DeepSeek's API capabilities in line with its App and web versions. Previously, developers often encountered issues with insufficient context when processing long texts through the API, requiring manual segmentation, while web users already enjoyed million-level context.

The expansion of the context window means a significant increase in the amount of text that can be processed in a single conversation—128k Tokens is approximately equivalent to a short story, and 1M Tokens can accommodate an entire set of "Dream of the Red Chamber." For developers who need to process complete contracts, research reports, or codebases, this change directly eliminates the main barrier to use.

The update to the knowledge base is also noteworthy. Developers have reported that, without an internet connection, the model can accurately answer news events from April 2025, whereas previously the model's knowledge cutoff was停留在2024.

There is a clear limitation to this update: it does not support visual input, only processing text and voice, and does not have multimodal capabilities.

Some developers believe that the version running behind this API may be the lightweight version V4 Lite of DeepSeek V4. According to various sources, DeepSeek founder Liam Liang has disclosed in internal communications that V4 will be officially released in late April, and the characteristics of this API update—long context, quick inference, and no support for multimodality—align with the positioning of the lightweight preview version.

However, DeepSeek has not made any explanation regarding the changes to the API model, and the latest official announcement on its website still pertains to the update content of DeepSeek-V3.2, where the context window remains at 128k.