MiniMax M2.7: AI Achieves Self-Evolution, Taking a Critical Step Toward AGI
Released on March 18, 2026, MiniMax M2.7 marks the next generation of large language models from the brand, coming just one month after the launch of its predecessor, M2.5. This is no ordinary upgrade of model parameters or a refresh of benchmark rankings; it represents a milestone breakthrough in the evolution of artificial intelligence – M2.7 has become the world’s first AI model to deeply participate in its own iterative optimization. As AI begins to rewrite its own code and lead the training and optimization process like an engineer, we can’t help but wonder: is the distance between artificial intelligence (as a tool) and Artificial General Intelligence (AGI) rapidly closing?
In the past, the iteration and upgrade of AI models have always relied on the full leadership of human engineers, with models merely acting as passive tools to execute instructions. The advent of MiniMax M2.7 has seen AI shift from “passive execution” to “active evolution” for the first time. It can undertake 30% to 50% of work tasks in specific R&D scenarios, with human researchers only intervening to control key decision-making nodes. This shift in model validates the “Silicon Valley Consensus” proposed by Eric Schmidt, former CEO of Google: as AI’s reasoning capabilities and memory systems continue to develop, it will reshape human work and operational methods, ultimately achieving recursive self-improvement and learning and evolving at a speed incomprehensible to humans.
From a Passive Tool to Active Evolution: Core Technological Breakthroughs of M2.7
The core of MiniMax M2.7’s realization of AI self-evolution lies in the construction of the AgentHarness framework, which enables the model to deeply participate in the entire process of its own training, evaluation and optimization. Instead of simply executing human-given instructions, it possesses the abilities of independent planning, autonomous execution and self-verification. To demonstrate this self-evolution capability in a more intuitive way, we can look at its real performance in two practical application cases within MiniMax.
Case 1: Undertaking 30%-50% of Core Work in Reinforcement Learning R&D
In MiniMax’s core internal RL (Reinforcement Learning) experimental workflow, M2.7 can independently perform 30% to 50% of the work content. It is capable of building complex Skills systems on its own and driving the design of its training process, model evaluation and effect optimization based on experimental objectives, without the need for detailed guidance from human engineers throughout the process.
This capability has directly led to a significant improvement in R&D efficiency, not only shortening the iteration cycle of reinforcement learning experiments remarkably but also allowing the model to identify problems in the experimental process more quickly and adjust directions independently. It frees human researchers from tedious basic work, enabling them to focus on more core strategy formulation and decision-making links.
Case 2: Completing Over 100 Iterations Independently, Boosting Evaluation Performance by 30%
In the optimization task of internal software engineering scaffolding, M2.7 has demonstrated an even more extreme self-evolution capability – running independently without human intervention throughout the process and completing more than 100 closed-loop cycles of “analysis-improvement-verification”.
In this process, the model is not just conducting simple parameter tuning, but truly understanding the optimization objectives of the software engineering scaffolding. Through continuous trial and error and verification, it independently discovers the optimal combination of sampling parameters and designs targeted workflow optimization solutions. Ultimately, this independent optimization has boosted MiniMax’s internal evaluation performance by 30%. This ability to understand objectives and find optimal solutions is already very close to the working thinking and methods of human engineers.
Robust Engineering Capabilities: From “Writing Code” to “Delivering Products”
The practical value of an AI model is measured not only by its theoretical evolutionary capabilities but also by its actual engineering implementation capabilities. MiniMax M2.7 has achieved top results in a number of professional and rigorous benchmark tests, proving that it is not just a model with chat interaction capabilities, but an artificial intelligence system with real engineering delivery capabilities that can solve complex problems in actual production environments.
Key Benchmark Test Results at a Glance
To clearly demonstrate M2.7’s engineering capabilities, we have collated its key scores in core benchmark tests below. These tests are all targeted at real engineering scenarios and can objectively reflect the model’s practical application capabilities:
From the test results, it can be seen that M2.7’s score in SWE-Pro is close to the best level of Claude Opus, and it ranks in the first echelon of the industry in all three core engineering tests, fully proving the professionalism and practicality of its engineering capabilities.
Code Repair: Resolving Senior Engineer-Level Online Faults in 3 Minutes
The excellent results achieved in the SWE-Pro test have been directly translated into M2.7’s code repair capabilities in actual production environments. In real online business scenarios, M2.7 can realize the full-process operation of automatically associating monitoring data, accurately locating the root cause of Bugs, and independently writing repair scripts, reducing the recovery time of online faults that originally required senior engineers to handle to less than 3 minutes.
This capability is of great significance for enterprise technical operation and maintenance. The resolution efficiency of online faults is directly related to the normal operation of businesses. The involvement of M2.7 can significantly reduce business losses caused by faults and improve the operation and maintenance efficiency of technical teams.
End-to-End Project Delivery: Full Process Control from Requirement Analysis to Product Launch
M2.7’s outstanding performance in the VIBE-Pro end-to-end project task test confirms that it has the complete project delivery capability from requirement analysis, solution design, code development to product launch. This means that M2.7 is no longer just a “code tool” that can complete code completion and snippet writing, but a “project engineer” that can understand the overall project objectives and control the entire project process.
For enterprise R&D teams, this capability can effectively assist in completing the full process from requirements to implementation. Especially in small and medium-sized R&D teams, it can effectively make up for the shortage of human resources and improve the overall project R&D efficiency.
Complex System Understanding: Mastering Terminal-Level Complex Architectures in Depth
Terminal Bench 2 is a test with extremely high requirements for complex system understanding capabilities. M2.7’s score of 57.0% in this test proves that it has the ability to deeply understand and operate terminal-level complex system architectures. In practical applications, this capability allows M2.7 to handle various complex system-level tasks, whether it is system architecture design, complex process debugging, or terminal system optimization and upgrading, and it can provide professional solutions.
Native Multi-Agent Collaboration: Completing Complex Tasks Through Division of Labor and Cooperation Like a Human Team
Another core capability of MiniMax M2.7 is its native Agent Teams (multi-agent collaboration) capability. This capability allows the AI model to no longer be an individual working “alone”, but to carry out division of labor and cooperation independently, and complete long-chain, high-complexity work tasks without the support of complex external frameworks.
In complex software engineering scenarios, the value of this capability is particularly prominent. Just like a professional human development team, where front-end engineers, back-end engineers and test engineers perform their respective duties and cooperate with each other, M2.7 can independently coordinate multiple agents, assign tasks matching their capabilities to different agents, and ultimately complete the full process from product design, code development to functional testing. This multi-agent collaboration model enables AI to handle more complex and systematic engineering tasks, further approaching human working methods.
Empowered by OpenClaw: M2.7 Transforms from a Technical Model to a Practical “Digital Employee”
If the technological breakthroughs of MiniMax M2.7 let us see the possibility of AI self-evolution, then its combination with OpenClaw has made this advanced AI capability truly integrated into the daily workflow of ordinary people, transforming M2.7 from a technical model in the laboratory into a practical “digital employee” available 24/7.
OpenClaw: An Open-Source AI Automation Assistant with 100,000+ Stars in 2 Days
OpenClaw is an open-source personal AI automation assistant that became a hit on the GitHub platform in early 2026. Relying on its strong adaptability and practicality, it gained more than 100,000 stars just 2 days after its launch, becoming a popular tool in the developer community. Its core value lies in building a universal bridge for the implementation of AI capabilities: it can connect to various large models (including MiniMax M2.7) at the upper level and adapt to various terminal platforms such as Feishu and Telegram at the lower level, allowing users to directly call the capabilities of top large models in the office and social tools they use daily to build their own cloud AI assistants.
To make it more convenient for ordinary users to use this tool, MiniMax officially launched MaxClaw, which migrates OpenClaw to the web end and realizes “one-click deployment and out-of-the-box use”. Users do not need to have professional development capabilities to quickly experience the powerful capabilities of the combination of M2.7 and OpenClaw.
Real Application Case: Building a Custom “Digital Avatar” Robot
A developer successfully built a custom “digital avatar” robot through the combination of OpenClaw + MiniMax M2.7 + ActivityWatch. This intelligent agent can deeply analyze the user’s work and usage habits and become a “digital companion” that truly understands the user. Its core functions include:
-
App usage distribution analysis: Automatically count the duration and frequency of the user’s daily use of various Apps, clearly showing which Apps are the core of use; -
Time allocation analysis: Precisely distinguish the proportion of time spent in different scenarios such as work, entertainment and rest, allowing users to have an intuitive understanding of their time usage; -
Efficiency analysis: Identify the time period with the highest work efficiency for the user through the correlation analysis of the user’s work results and time input; -
Personalized improvement suggestions: Based on the above analysis results and combined with the user’s work objectives, provide targeted suggestions for time management and efficiency improvement; -
Automatic report push: Without manual operation by the user, generate an analysis report automatically every day and push it to the user to help the user complete the daily work review.
This “digital avatar” is not just a simple data analysis tool, but an intelligent assistant that can understand the user’s work needs and fit the user’s usage habits, truly realizing the personalized AI assistance for personal work.
Feishu/Telegram Access: Building a 24/7 Personal “Super Brain”
After connecting M2.7 to daily office and social terminals such as Feishu and Telegram through OpenClaw, with the help of M2.7’s long-term memory framework, this intelligent agent can become the user’s personal “super brain”, realizing a number of practical functions covering multiple scenarios of work and life:
-
Immersive role-playing: Can integrate real emotions and role settings to become the user’s personal assistant, psychological counselor, or even virtual companion, meeting different emotional and interactive needs; -
Complex Office automation: Different from simple text generation by ordinary AI, it can handle various high-level Office tasks such as complex Excel data analysis, long Word document editing, and professional PPT design, truly replacing manual work to complete office tasks; -
Scheduled tasks and web search: Realize 24/7 online operation, automatically execute scheduled tasks according to user settings, and support real-time web search to obtain the latest information and data to support user decision-making.
Zero-Cost Efficient Programming: Individual Developers Can Also Access the Capabilities of Top Models
For developers, the MiniMax M2.5/M2.7 Free API provided by the Zen platform enables zero-cost efficient programming with OpenClaw. Developers can configure a dual-model switching strategy in OpenClaw and call the matching model according to different task types, balancing development efficiency and usage costs:
-
Core programming tasks: Call MiniMax M2.7 to leverage its top-notch code understanding, development and debugging capabilities to complete complex programming work; -
Result summary and display: Call other lightweight models to complete work such as code comments, result sorting and report generation, reducing the overall API call cost.
This dual-model combination allows individual developers to access the capabilities of top large models without bearing high model usage costs, realize complex agent collaboration development, and significantly lower the threshold for AI development.
Professional Office Field: A Qualitative Leap from “Text Generation” to “In-Depth Understanding”
In the core office scenarios of enterprises and individuals, MiniMax M2.7 has also demonstrated capabilities far exceeding those of traditional AI models. Its core transformation is an upgrade from simple “text generation” to “in-depth understanding” of office needs, enabling it to truly handle complex office tasks and provide unexpected solutions.
In the GDPval-AA evaluation covering professional knowledge in multiple fields, M2.7 achieved an ELO score of 1495. This result proves that it has a solid reserve of professional knowledge in various fields and can meet the professional office needs of different industries. In actual office tests, M2.7’s performance in handling tasks related to office software such as Excel, Word and PPT has been described as “extremely smooth”, which is specifically reflected in three aspects:
-
Accurate demand understanding: Can accurately capture the user’s core office needs, and can extract key information through in-depth analysis even for vague demand descriptions; -
Independent output optimization: Instead of simply completing tasks according to user instructions, it will independently optimize the output results according to the professional requirements of office scenarios, such as automatically selecting more appropriate charts in Excel analysis, optimizing the logical flow of writing in Word editing, and matching layouts more suitable for the theme in PPT design; -
Providing solutions: For complex office tasks, it can not only complete the basic execution work but also provide a variety of solutions for the user and analyze the advantages and disadvantages of different solutions to help the user make a more appropriate choice.
Whether it is the daily office collaboration of enterprises or the efficient office needs of individuals, M2.7 can become a professional office assistant tool, greatly improving office efficiency and reducing the time and energy costs of office work.
Recognition from the Global Developer Community: The Top Strength of Domestic Models
Since its release, MiniMax M2.7 has aroused a huge response in the global developer community. Its performance in various international evaluation rankings and usage data has fully proved the top strength of domestic large models, and also let the industry see that the gap between open-source and closed-source models is narrowing rapidly.
Evaluation Rankings: 4th in the World and 1st Among Domestic Models on PinchBench
On the internationally renowned PinchBench ranking, M2.7 has achieved the result of 4th in the world and 1st among domestic models. This ranking confirms its top position among global large models and also allows domestic large models to occupy an important place on the international stage.
Usage Data: Topping the Charts for Four Consecutive Weeks on OpenRouter
On the OpenRouter platform, the annualized throughput of MiniMax M2.5’s global model tokens has exceeded 1 quadrillion, and it has topped the charts for four consecutive weeks with stable performance and strong capabilities, becoming one of the most popular large models among developers on the platform.
Developer Feedback: A Cost-Effective Choice for Top Models
The launch of M2.7 has provided global developers with a cost-effective choice of AI models. A developer bluntly said: “Building an agent with MiniMax M2.7 only costs 1/10 of the price of Claude Opus – it’s an amazing deal!” An industry practitioner also sighed: “The gap between open-source and closed-source models is narrowing every month; cutting-edge laboratories should be alert.”
These feedbacks are not only a recognition of MiniMax M2.7’s capabilities but also reflect that domestic large models are occupying an increasingly important position in the global artificial intelligence field with their rapid iteration speed, providing more high-quality choices for global developers.
How Close Are We to AGI?
When MiniMax M2.7 can independently complete more than 100 rounds of iterative optimization, when it can repair online faults at the senior engineer level in 3 minutes, and when it can complete complex project delivery through multi-agent collaboration like a human team – a frequently asked question is once again put in front of us: How close are we to Artificial General Intelligence (AGI)?
The prediction by Eric Schmidt, former CEO of Google, is gradually becoming a reality: AI systems are developing toward recursive self-improvement and may learn and evolve at a speed incomprehensible to humans in the future. The emergence of MiniMax M2.7 has let us see the real embryonic form of AI self-evolution. It is no longer just a tool to execute instructions, but has the abilities of independent thinking, autonomous execution and self-optimization – these abilities are the core foundation for moving toward AGI.
But what is more noteworthy is that M2.7 is not a technical achievement confined to the laboratory. Its combination with OpenClaw has made AI’s self-evolution capability truly integrated into the workflow of ordinary people, allowing every user to experience the practical value of AI as a “digital employee”. It is no longer just a cold dialogue box, but an intelligent companion that is online 24/7, has long-term memory, can self-evolve and can truly solve practical problems.
Perhaps the arrival of Artificial General Intelligence (AGI) will not be a sudden “singularity”, nor the sudden birth of a certain model on a certain day. Instead, like MiniMax M2.7, through continuous technological breakthroughs, artificial intelligence is gradually transformed from a “passive tool” to an “active assistant”, and from an “intelligent companion” to a general artificial intelligence with autonomous consciousness step by step.
The release of MiniMax M2.7 is only a critical step for artificial intelligence toward AGI, but the model transformation brought about by this step has let us see the future development direction of artificial intelligence. When AI can continuously achieve self-evolution and this evolutionary capability can be truly applied in practice, the arrival of general artificial intelligence may be closer than we think.
FAQ: Key Questions Answered About MiniMax M2.7
1. When was MiniMax M2.7 released?
MiniMax M2.7 was officially released on March 18, 2026, just one month after the launch of its predecessor M2.5, marking a rapid and milestone model iteration.
2. What is the core breakthrough of MiniMax M2.7?
M2.7’s core breakthrough is becoming the first AI model to deeply participate in its own iteration. The AgentHarness framework it built enables the model to deeply participate in its own training, evaluation and optimization processes, and can undertake 30% to 50% of work tasks in specific R&D scenarios.
3. What are the key capabilities of MiniMax M2.7?
M2.7’s core key capabilities include: the AgentHarness framework, native multi-agent collaboration capabilities, top-tier engineering delivery capabilities, in-depth professional knowledge understanding capabilities, and autonomous self-iterative optimization capabilities.
4. What are MiniMax M2.7’s core results in benchmark tests?
In the three core engineering benchmark tests, M2.7’s scores are: 56.22% in SWE-Pro, 55.6% in VIBE-Pro and 57.0% in Terminal Bench 2, all ranking in the first echelon of the industry; it achieved an ELO score of 1495 in the GDPval-AA professional evaluation and ranked 4th in the world and 1st among domestic models on the PinchBench ranking.
5. What are the main application scenarios of MiniMax M2.7?
M2.7’s application scenarios cover various aspects of technological R&D and daily office work, with core scenarios including: code repair and development, end-to-end project delivery, complex system architecture processing, multi-agent collaboration engineering tasks, complex Office automation, personal AI assistant, and digital avatar creation.
6. How to experience the capabilities of MiniMax M2.7?
There are currently three main experience entrances to meet the needs of different users:
-
MiniMax Agent: Experience the core agent capabilities of the model directly; -
MaxClaw: The web version of OpenClaw officially launched by MiniMax, with one-click deployment and out-of-the-box use; -
Zen Platform: Provides free MiniMax M2.5/M2.7 API keys for zero-cost experience of model capabilities.
7. What is the value of the combination of OpenClaw and MiniMax M2.7?
As an open-source AI automation assistant, OpenClaw builds a bridge between large models and terminal platforms, enabling the advanced capabilities of M2.7 to be implemented in daily terminals such as Feishu and Telegram. It transforms M2.7 from a laboratory model into a practical “digital employee” available 24/7, truly integrating it into the user’s daily workflow.
8. What are the characteristics of MiniMax M2.7’s multi-agent collaboration capability?
M2.7 has native Agent Teams multi-agent collaboration capabilities. Without the support of complex external frameworks, it can independently assign work and cooperate among multiple agents, completing long-chain and complex engineering tasks from design to testing like a human development team. This is an important feature that distinguishes it from other models.
