陈天桥代季峰打响2026大模型第一枪：30B参数跑出1T性能

Core Viewpoint - MiroThinker 1.5, developed by MiroMind, is positioned as a leading AI model in the intelligent agent field, showcasing superior performance in various benchmark tests compared to other top models like GPT-5-High and Gemini-3-Pro [1][3][5]. Performance Evaluation - MiroThinker 1.5 achieved notable scores in benchmark tests: - HLE-Text: 39.2% - BrowseComp: 69.8% - BrowseComp-ZH: 71.5% - GAIA-Val-165: 80.8% [3][4]. - It surpassed ChatGPT-Agent's previous record in BrowseComp, establishing itself in the global top tier [5]. Model Efficiency - MiroThinker 1.5 operates with significantly fewer parameters (30B and 235B) compared to competitors, achieving comparable or superior results through high efficiency [7][8]. - The model's inference cost is notably low at $0.07 per call, which is only 1/20 of Kimi-K2-Thinking's cost, while also demonstrating faster inference speeds [8]. Development Team and Background - The MiroMind team, responsible for MiroThinker 1.5, previously excelled in predicting outcomes in decentralized markets, showcasing their expertise in model development [9][10]. Interactive Scaling and Model Training - MiroThinker 1.5 incorporates a novel approach called Interactive Scaling, which emphasizes interaction with the external environment during both training and inference phases, enhancing its reasoning capabilities [46][58]. - The model employs a feedback loop in its reasoning process, allowing for iterative verification and correction, which contrasts with traditional models that rely heavily on memorization [48][57]. Predictive Capabilities - MiroThinker 1.5 demonstrates a robust ability to make predictions based on real-time data, as evidenced by its analysis of sports events and video game release timelines, showcasing a logical and evidence-based approach [15][35][41]. - The model's predictions are structured to avoid reliance on past knowledge, instead focusing on current information and real-world interactions [52][63]. Conclusion - MiroThinker 1.5 represents a significant advancement in AI model development, prioritizing interaction and evidence-based reasoning over sheer parameter size, thus redefining the landscape of intelligent agents [64].