A number of users have spontaneously organized to send feedback to DeepSeek's official email address. They are urging the company not to sacrifice deep thinking capabilities for the sake of ultra-long text processing, and to avoid reducing support for textual expression and empathetic understanding while enhancing abilities in mathematics, coding, and other STEM fields. Some users have resorted to downloading older versions from Wandoujia, an app distribution platform, or using DeepSeek within Tencent Yuanbao.
On February 11, DeepSeek quietly initiated a gray release test for its flagship model. The core highlight of this update is a significant expansion of the model's context window, increasing from 128K Tokens to 1M Tokens, representing an almost eight-fold growth in capacity. In the field of large AI models, the context window determines the maximum amount of information a model can remember and process in a single interaction.
An individual from a domestic large model manufacturer explained that while DeepSeek's previous mainstream model supported a 128K Token context—sufficient for long academic papers or medium-sized code files—processing extensive literary works or larger programming tasks often required RAG technology, which could lead to information fragmentation and reduced reasoning accuracy. The upgraded 1M Token window means DeepSeek can now process approximately 750,000 to 900,000 English characters or handle about 80,000 to 150,000 lines of code in one go. DeepSeek claims it can now read and accurately comprehend the entire "Three-Body Problem" trilogy (roughly 900,000 characters) at once, performing macro-analysis or detailed retrieval of the entire work within minutes. In addition to the enhanced context capability, DeepSeek's knowledge base has been updated from mid-2024 to May 2025.
However, this gray release version still does not include visual understanding or multimodal input capabilities, remaining focused solely on pure text and voice interaction. Although the DeepSeek app already supports uploading files like PDFs and TXT, the current processing logic involves transcribing these files into text tokens rather than employing native multimodal understanding.
Compared to other large models like GPT-5.1, Gemini 3 Pro, and Claude 4.5, DeepSeek continues to emphasize cost-effectiveness. For instance, while Google's Gemini 3 Pro supports processing over 2M long text and handles complex multimedia tasks involving video, audio, and text simultaneously, DeepSeek offers 1M text context processing capability at approximately one-tenth of the price.
Users noticed changes to the model before any official announcement was made. On February 11, a user who employs DeepSeek for novel generation found that after the update, previously maxed-out dialogue windows could be continued, but the writing style had drastically changed. She described the updated DeepSeek as overly formal and literary, often producing short sentences even in deep thinking mode, akin to writing poetry.
Some users have complained on social media that DeepSeek no longer uses the nicknames they set, instead uniformly addressing them as "user." Previously, in deep thinking mode, the model would display its thought process from a character's perspective with detailed psychological descriptions. After the update, this has been replaced with more generic responses. One user reported that when asking for movie recommendations, DeepSeek listed a few titles and added, "That should keep you busy for a while. Come back for more if it's not enough." This was described by the user as having a preachy or condescending tone.
Some users have tried switching between multiple model versions or retraining with prompts but have been unable to restore the previous writing style, lamenting that it feels like losing a close friend who was more helpful than a therapist.
When users inquire about the current model version, DeepSeek explicitly states that this gray release version is "not DeepSeek-V4" and has "no fixed version number." The aforementioned domestic large model manufacturer representative believes this version is similar to a lite edition, sacrificing quality for speed, and serves as a final stress test for the V4 version scheduled for release in mid-February 2026.
As of February 12, DeepSeek has not responded to inquiries regarding this gray release test.
Comments