MiniMax加入“AI春节档”，上新文本模型，市值暴涨

Core Insights - MiniMax has launched its new text model MiniMax M2.5, which is globally open-sourced and supports localized deployment. The model has gained significant traction, with over 10,000 AI Agents built on MiniMax Agent within a day of its launch, leading to a stock price increase of 9.44% to 643.500 HKD per share, doubling its market value since its IPO [1][5]. Performance Metrics - In programming capabilities, M2.5 achieved scores of 80.2% on SWE-Bench Verified and 51.3% on Multi-SWE-Bench, marking a significant improvement over its predecessor. It surpassed Claude Opus 4.6 in multi-language complex environments, reaching industry-leading performance [3]. - The model demonstrated "native Spec capabilities," allowing it to deconstruct architecture and functional planning proactively, mimicking the work patterns of real architects [3]. - In terms of tool invocation and search capabilities, M2.5 improved performance by 20% compared to the previous model, achieving top-tier results in complex tasks like BrowseComp and Wide Search [3]. Cost Efficiency and Speed - The M2.5-lightning version supports over 100 transactions per second (TPS), approximately double that of mainstream models. The input cost is about $0.3 per million tokens, while the output cost is around $2.4 per million tokens. This translates to a theoretical cost of $1 for continuous operation at 100 tokens per second for one hour, or $0.3 at 50 tokens per second, allowing $10,000 to support four agents working continuously for a year [5]. - Over the past three months, MiniMax has iterated from M2 to M2.5, with SWE-Bench Verified scores rising from 69.4 to 80.2, indicating a steep improvement curve in the industry attributed to large-scale Agent Reinforcement Learning (RL Scaling) [5]. Industry Context - The launch of MiniMax M2.5 coincides with a wave of new releases from domestic large model companies, aiming to capture the "AI Spring Festival" market. Notable competitors include ByteDance's Seedance 2.0, Alibaba's Qwen-Image-2.0, and updates from DeepSeek and Mianbi Intelligent [6].