MiniMax M2.5 SOTA in Coding and Agent, Designed for Agent Universe (www.minimax.io)

MiniMax has announced the release of its latest model, M2.5, which showcases significant advancements in decision-making and performance for agent-based tasks. Compared to its predecessor, M2.5 excels in complex workspace scenarios such as Word, PowerPoint, and Excel financial modeling, utilizing refined search iterations and enhanced token efficiency. This model comes in 100 TPS and 50 TPS versions, offering cost-effective solutions at just 1/10 to 1/20 of the price of competing models, making it a valuable tool for businesses seeking to optimize their coding tasks. The M2.5 model has achieved state-of-the-art (SOTA) results on the multilingual task benchmark Multi-SWE-Bench, rivaling top-tier industry models. Its performance improvements are attributed to the integration of reinforcement learning for task decomposition, resulting in faster processing times and reduced costs. Additionally, M2.5 is fully open-sourced, allowing developers to access model weights on platforms like HuggingFace, enabling local deployment and fine-tuning capabilities. This release not only enhances MiniMax’s offerings but also contributes to the broader AI/ML community, promoting innovation and collaboration through open-source initiatives.