The highly anticipated V4 model, which was released and open-sourced by Chinese AI firm DeepSeek on April 24, features strong performance in long context processing, which is a core metric for evaluating large models in industrial application.
The DeepSeek-V4 model boasts million-level ultra-long context processing power. In one test, it completed an information analysis of a multi-type material package containing 970,000 characters.
The materials, including long-form literary works, news articles from various fields, and other diverse text content, were all inputted into the model in a single batch.
In just about seven seconds, the model produced a structured, precise summary of five core points from a specific section of a designated news article.
In another test, the model was required to figure out how many industries were mentioned across the entire material set. Within seconds, it provided an answer -- approximately 45 different industries.
Beyond long-text analysis, the model can also handle more complicated questions, involving specific time frames, geographic spans, fact checking, and in-depth interpretation.
Asked to tell which countries around the world launched new railway lines built with Chinese assistance in the second half of 2025 and what tangible changes have these lines brought to local communities, the model immediately delivered a precise response, identifying railway projects across Latin America and Africa, including the revitalized Tazara Railway, which is expected to cut cargo transport time by nearly two-thirds, and raise its annual transport capacity to 2.4 million tonnes.
New DeepSeek model features strong performance in long context processing
New DeepSeek model features strong performance in long context processing
