刚刚！DeepSeek，重大发布！

Core Viewpoint - The article highlights the recent advancements in domestic AI models, particularly focusing on the release of new models by DeepSeek, Kimi, and Alibaba, which demonstrate significant improvements in performance and capabilities in visual understanding and reasoning [2][8][9]. Group 1: DeepSeek's Innovations - DeepSeek released the new DeepSeek-OCR 2 model on January 27, utilizing the innovative DeepEncoder V2 method, allowing AI to dynamically rearrange image components based on their meaning, mimicking human visual encoding logic [2][3]. - The DeepSeek-OCR 2 model achieved a score of 91.09% on the OmniDocBench v1.5 benchmark, representing a 3.73% improvement over its predecessor [3][4]. - The model maintains high precision while controlling computational costs, with visual token counts limited between 256 and 1120, aligning with Google’s Gemini-3 Pro [3][4]. Group 2: Kimi's Model Release - Kimi launched the Kimi K2.5 model, which automatically updated the previous K2 model without user intervention, aimed at enhancing response speed, reasoning ability, and multi-turn dialogue stability [8]. - Kimi K2.5 achieved top scores in various agent evaluations, including HLE, BrowseComp, and DeepSearchQA, marking it as Kimi's most intelligent model to date [8]. Group 3: Alibaba's Qwen3-Max-Thinking - Alibaba introduced the Qwen3-Max-Thinking model, which surpassed leading models like GPT-5.2 and Claude Opus 4.5 in multiple key performance benchmarks, setting new global records [8][9]. - The new model features over one trillion parameters and incorporates a novel test-time scaling mechanism, significantly enhancing reasoning performance while being more economical [9]. - Qwen3-Max-Thinking also improves the model's ability to autonomously use tools, reducing hallucinations and laying the groundwork for solving complex real-world tasks [9].