DeepSeek's V4 Programming Skills Spark Online Buzz and Style Trend Discussions

On February 11, DeepSeek started rolling out a gradual update to its web and mobile app platforms. By February 14, the company officially announced this as a trial of its new long-text model architecture. The updated web and app interfaces now handle ultra-long contexts of up to 1 million tokens, with the knowledge base refreshed to include information through May 2025. However, the API service remains on version V3.2, which still supports only 128K context. Industry experts view this update as a technical warm-up and stress test ahead of the launch of its next-generation V4 model. Consequently, online anticipation is building across platforms like Weibo for the upcoming V4 release.

Following the update, DeepSeek's interaction style shifted noticeably, prompting user feedback that it "became less personable." This topic quickly rose to the top of Weibo's trending list, amassing over 68.535 million views. Specific changes include the model no longer using personalized nicknames for users, instead addressing everyone uniformly as "user." In deep thinking mode, responses became shorter and more direct, adopting a straightforward and concise tone. Some replies were even perceived as "ambiguous in tone" or "sarcastic," leaving users accustomed to its previous empathetic style feeling a sense of withdrawal. Netizens' reactions were split: emotional fans missed the earlier warm interactions, while efficiency-minded users praised the rational and streamlined communication, seeing it as the hallmark of a true productivity tool. Notably, the updated model successfully passed the "car wash Turing test," a challenge that often trips up leading AI models.

Regarding the style shift, DeepSeek's official statement clarified that it was not an intentional change but a byproduct of prioritizing efficiency and refining response boundaries. Excessive expressions and filler content could dilute information density in complex queries, and the update also catered to users who prefer clear answers without what they see as "AI pretending to care." This development has further fueled online excitement for DeepSeek V4. Reports suggest the model is slated for release around the Spring Festival in mid-February 2026, possibly on or near February 17, according to Weibo sources.

As a flagship model with trillions of parameters, V4 places a strong emphasis on programming enhancements. Early internal benchmarks indicate that V4 has already outperformed leading models like Claude and GPT in programming tasks, positioning it to potentially reshape the AI programming landscape. Additionally, V4 introduces several key technological advances, allowing it to process and interpret extremely long code prompts and comprehend large codebase contexts in a single pass—a significant advantage for enterprise-level development. Upgraded training algorithms enhance its ability to recognize data patterns and reduce performance degradation. The model's reasoning is more rigorous and dependable, achieving stable performance without regression while striking a better balance across various capabilities.

At the same time, the model will maintain its million-token context advantage, with inference costs significantly lower than those of Western competitors. It is also scheduled to be open-sourced under the Apache 2.0 license.