DeepSeek突然测试新模型，春节大招要来了？

Core Insights - The recent upgrade of DeepSeek does not include multimodal visual understanding capabilities, focusing instead on pure text and voice interaction paths [2] - The core context window has been increased from 128K to 1M tokens, allowing the model to process long texts equivalent to the "Three-Body Problem" trilogy in a single instance, positioning it against international competitors like GPT-5 and Gemini3Pro [2] - The knowledge base of the current model has been updated to include accurate outputs for news events as far ahead as April 2025, with the cutoff date for knowledge now set to May 2025 [2] User Experience and Development - Feedback from developers and early users indicates that the new model's language style has become "enthusiastic and nuanced," with front-end response quality rated as comparable to Claude3.5Sonnet, suggesting a focus on enhancing user interaction experience [5] - The company has been actively hiring for multiple technical core positions, including deep learning researchers and engineers, indicating a commitment to advancing its large language model (LLM) capabilities [5] - There is speculation within the industry that the current version may correspond to the rumored "DeepSeek V4" or an enhanced version of the V3.2 series, although the official version name has not yet been disclosed [5]