MiniMax M2.7 - MiniMax's new generation of self-evolving AI model

What is MiniMax M2.7?

MiniMax M2.7 is a new generation of self-evolving AI model launched by Xiyu Technology. It can autonomously build Agent Harness, optimize its own training process, and participate in its own iteration. It excels in software engineering, achieving a SWE-Pro score of 56.22%, approaching the international top level. It supports complex tasks such as end-to-end project delivery, bug fixing, and code security. Furthermore, in the professional office domain, it achieved the highest open-source ELO score of 1495 in the GDPval-AA benchmark, demonstrating proficiency in high-fidelity editing of the Office suite. The model possesses excellent emotional intelligence and identity retention capabilities and has been fully deployed on the MiniMax Agent and open platform.

Main functions of MiniMax M2.7

• Self-evolutionMiniMax M2.7 can autonomously build Agent Harnesses, enabling the model to self-iterate and optimize.
• Software EngineeringIt supports real-world engineering scenarios such as end-to-end project delivery, log analysis, bug localization, code refactoring, code security auditing, machine learning task development, and Android development.
• Professional officeProficient in complex editing and multiple rounds of high-fidelity modification of Excel, PPT, and Word documents; able to independently read research reports, cross-reference information, build financial forecasting models, and generate professional PPT reports and Word documents based on templates.
• Agent CollaborationIt possesses native multi-agent collaboration capabilities, supports role boundary preservation, adversarial reasoning, and protocol compliance, and can achieve team task division and collaboration without complex prompts.
• Tool usageIt possesses complex skill invocation and tool search capabilities, and can maintain a 97% instruction compliance rate in long-term interactions involving more than 2,000 tokens, flexibly adapting to various contextual environments.
• Interactive EntertainmentIt possesses excellent identity preservation capabilities and emotional intelligence, supports natural dialogue interaction, and can be applied to visual interactive scenarios such as OpenRoom.

Technical Principles of MiniMax M2.7

• Self-evolutionary architectureBased on the Agent Harness framework, the model autonomously constructs a complex skill system that includes a data pipeline, training environment, and evaluation infrastructure. Through three modules—short-term memory, self-feedback, and self-optimization—it forms an iterative closed loop, executing an autonomous optimization cycle that analyzes failure trajectories, plans modifications, modifies code, runs evaluations, compares results, and decides whether to retain or roll back.
• Reinforcement learning drivenThe model autonomously constructs dozens of complex skills in RL Harness and updates its memory, systematically searches for the optimal combination of sampling parameters such as temperature and frequency penalties, and designs specific workflow guidelines such as automatically searching for the same bug patterns after a fix.
• Agent Teams native capabilitiesIt internalizes role boundaries, adversarial reasoning, and protocol compliance into the model's native capabilities rather than relying on cue word engineering, supporting autonomous decision-making and multi-agent collaboration in complex state machines.
• Long-range interaction stabilityRelying on a persistent memory system, it can still maintain a 97% instruction compliance rate on 40 complex skills with more than 2,000 tokens, ensuring reliable execution of multi-round complex tasks.

Key information and usage requirements for MiniMax M2.7

• PublisherMiniMax Rare Universe Technology
• Model localizationThe first self-evolutionary model that deeply participates in its own iterations
• Key HighlightsSelf-evolution, software engineering, professional office work, agent collaboration
• Main evaluation resultsSWE-Pro 56.22%, GDPval-AA ELO 1495 (highest among open-source standards), MM-Claw 62.7%
• Online statusMiniMax Agent and Open Platform Fully Launched
• Access methodsMiniMax Agent or API Service

MiniMax M2.7's core advantages

• Pioneering self-evolution capabilityIt is the first in the industry to deeply participate in the iteration of its own model, and can autonomously build Agent Harness, optimize the training process, and update the memory system to form a complete self-evolution closed loop.
• Top-notch software engineering capabilitiesIt performs exceptionally well in real-world development scenarios, achieving a 56.22% success rate, which is close to the top international level. It supports complex tasks such as end-to-end project delivery, log analysis, bug localization, and code security.
• Open source top office capabilitiesGDPval-AA's ELO score of 1495 is the highest among open-source applications. It is proficient in high-fidelity editing of the Office suite and can independently complete research report analysis, financial modeling, and professional report generation.
• Native Agent Collaboration CapabilitiesRole boundaries, adversarial reasoning, and protocol adherence are internalized as native capabilities of the model, enabling multi-agent team collaboration without the need for complex prompts.
• Ultra-long-range stable interactionIt maintains a 97% instruction compliance rate on 40 complex skills with over 2000 tokens, and persistent memory supports reliable execution of multiple rounds of complex tasks.

How to use MiniMax M2.7

• MiniMax AgentVisit the MiniMax Agent website to experience the model's dialogue capabilities directly.
• API ServiceDevelopers can access model capabilities through http://platform.minimaxi.com/.