MiniMax官宣参战“春节档” 新一代文本模型性能持续提升

Core Insights - MiniMax officially launched its new generation text model, MiniMax M2.5, on February 12, 2026, positioning it as a "native Agent production-level model" in the competitive landscape referred to as the "AI gods battle" [2] Group 1: Model Performance - In programming capabilities, MiniMax M2.5 achieved a score of 80.2% on the SWE-BenchVerified and 51.3% on the Multi-SWE-Bench, showing significant improvement over its predecessor [2] - The model surpassed Opus 4.6 in multi-language complex environments, reaching the industry's best performance [2] - The model exhibits "native Spec capabilities," actively decomposing architecture and functional planning before coding, closely resembling the work patterns of real architects [2] Group 2: Tool Utilization and Search Capabilities - The model can automatically handle complex tasks, achieving better results with lower round consumption in tasks like BrowseComp and Wide Search, showing a 20% improvement over the previous generation [3] - In office scenarios, MiniMax M2.5 demonstrated significant capability enhancements in high-level tasks involving Word, PPT, and Excel financial modeling, achieving an average win rate of 59.0% in the GDPval-MM evaluation framework compared to mainstream models [3] Group 3: Ecosystem Development - MiniMax M2.5 was launched on MiniMax Agent on February 12 and globally open-sourced for localized deployment on February 13, with over 10,000 experts built by users worldwide within a day [3] - The company aims to build a sustainable and scalable Agent ecosystem, referred to as Agent Universe, to enhance model capabilities, generalization, and cost-effectiveness, facilitating the penetration of Agents into various aspects of work and life, from programming to entertainment [3]