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

MiniMax M2.7 is a new generation of self-evolving AI model launched by Xiyu Technology. It can independently build Agent Harness, optimize its own training process, and participate in iterating itself. It has outstanding performance in software engineering, with model SWE-Pro scoring 56.22% close to the top international level, supporting complex tasks such as end-to-end project delivery, bug troubleshooting, and code security. At the same time, it achieved the highest open source ELO score of 1495 in the GDPval-AA evaluation in the professional office field, and is proficient in high-fidelity editing of the three-piece Office suite. The model has excellent emotional intelligence and identity maintenance capabilities and has been fully launched. MiniMax Agentand open platforms.

Main features of MiniMax M2.7

• self evolution : MiniMax M2.7 can independently build Agent Harness to realize self-iteration and optimization of the model.
• software engineering : Support end-to-end project delivery, log analysis, bug location, code reconstruction, code security audit, machine learning task development and Android development and other real engineering scenarios.
• Professional office : Proficient in complex editing and multiple rounds of high-fidelity modifications in Excel, PPT, and Word. Can independently read research reports, cross-compare information, build financial forecast models, and generate professional PPT reports and Word documents based on templates.
• Agent collaboration : It has native multi-agent collaboration capabilities, supports role boundary maintenance, adversarial reasoning and protocol compliance, and can achieve team task division and collaboration without complex prompt words.
• Tool usage : Possessing complex Skill calling and Tool Search capabilities, it can still maintain a command compliance rate of 97% in long-distance interactions of more than 2,000 Tokens, and can flexibly adapt to various contexts.
• interactive entertainment : It has excellent identity maintenance ability and emotional intelligence, supports natural dialogue interaction, and can be applied to visual interaction scenarios such as OpenRoom.

Technical principles of MiniMax M2.7

• self-evolving architecture : Based on the Agent Harness framework, the model independently builds a complex skill system including data pipeline, training environment, and evaluation infrastructure. An iterative closed loop is formed through the three modules of short-term memory, self-feedback, and self-optimization, and an autonomous optimization cycle is executed that analyzes the failure trajectory, plans changes, modifies the code, runs the evaluation, compares the results, and decides to retain or roll back.
• Reinforcement learning driven : The model builds dozens of complex skills in RL Harness by itself and updates memory independently, systematically searches for the optimal combination of sampling parameters such as temperature and frequency penalty, and designs specific workflow guidelines such as automatically searching for the same bug pattern after repair.
• Agent Teams native capabilities : Internalize role boundaries, adversarial reasoning, and protocol compliance into model native capabilities instead of relying on prompt word engineering to support autonomous decision-making and multi-agent collaboration in complex state machines.
• Long-range interactive stability : Relying on the persistent memory system, it can still maintain an instruction compliance rate of 97% on 40 complex skills with more than 2,000 Tokens, ensuring reliable execution of multiple rounds of complex tasks.

MiniMax M2.7 key information and usage requirements

• publisher ：MiniMax Xiyu Technology
• Model positioning : The first self-evolution model that deeply participates in iterating itself
• Core Highlights : Self-evolution, software engineering, professional office, Agent collaboration
• Main evaluation results : SWE-Pro 56.22%, GDPval-AA ELO 1495 (the highest in open source), MM-Claw 62.7%
• Online status : MiniMax Agent and open platform are fully launched
• Access method ：MiniMax Agent or API service

Core advantages of MiniMax M2.7

• The first self-evolution ability : The industry’s first in-depth participation in iterating its own model, it can independently build Agent Harness, optimize the training process, and update the memory system to form a complete self-evolution closed loop.
• Top software engineering capabilities : Excellent performance in real development scenarios, with SWE-Pro reaching 56.22%, close to the top international level, supporting complex tasks such as end-to-end project delivery, log analysis, bug location, and code security.
• Open source’s highest office capabilities : The GDPval-AA evaluation ELO score of 1495 is the highest among open sources. It is proficient in high-fidelity editing of the three-piece Office suite, and can independently complete research report analysis, financial modeling, and professional report generation.
• Native Agent collaboration capabilities : Role boundaries, adversarial reasoning, and protocol compliance are internalized into the model’s native capabilities, enabling multi-agent team collaboration without the need for complex prompt words.
• Ultra-long-range stable interaction : Maintaining a 97% instruction compliance rate on 40 complex skills with over 2,000 Tokens, and persistent memory supports the reliable execution of multiple rounds of complex tasks.

How to use MiniMax M2.7

• API service : Developers can access model capabilities through http://platform.minimaxi.com/.

Comparison of similar competing products of MiniMax M2.7

Dimensions	MiniMax M2.7	Claude Opus 4.6	GPT-5.4
self evolution	✅ First of all, the model participates in its own iteration	❌ None	❌ None
SWE-Pro	56.22%	About 56%+	Undisclosed specific scores
GDPval-AA	1495 (highest for open source)	About 1500+ (closed source is the strongest)	About 1490
MM-Claw	62.7%	close to horizontal	Not explicitly evaluated
Open source attributes	Partially open source	Closed source	Closed source
Available domestically	✅ Direct access	⚠️Need an agent	⚠️Need an agent
Core advantages	Self-evolution + real engineering + cost-effectiveness	Comprehensive strongest + long text	General capabilities + rich ecosystem

Application scenarios of MiniMax M2.7

• software development : MiniMax M2.7 can independently complete the full-process software engineering tasks from requirement analysis to code delivery, including production environment troubleshooting and repair and mobile application construction.
• Professional office : MiniMax M2.7 is good at handling the high-fidelity editing work of the three-piece Office suite. It can independently read research reports and build financial forecast models, and finally generate professional data analysis reports and presentation documents.
• Smart collaboration : MiniMax M2.7 supports multi-agent team collaboration mode, which can realize role division, adversarial reasoning and protocol compliance in complex projects, and can complete team task delivery without manual orchestration.
• Tool automation : MiniMax M2.7 has powerful tool usage capabilities, can maintain stable command following during long-distance interactions, and automatically calls various Skills to complete cross-system data integration and information research tasks.
• interactive entertainment : MiniMax M2.7 has excellent identity preservation capabilities and emotional intelligence, supports immersive role-playing and natural dialogue interaction, and can realize real-time scene exploration in visualization spaces such as OpenRoom. ©