The world of large language models (LLMs) has shifted in recent months from simple text generation to the creation of complex systems. MiniMax's latest announcement about the release of the M2.7 model shows that the next step is now: agentic workflows. This concept means that AI does not function as a passive tool, but as an active worker who has a set of skills, memory, and the ability to use external software to achieve a goal.
Self-Study: A Model That Builds Its Own Skills
What distinguishes MiniMax M2.7 from competing models, such as versions from OpenAI or Anthropic, is its ability for self-evolution, or self-study. According to the official announcement MiniMax News, the M2.7 model actively participates in its own development. This is not just a marketing ploy; it is a technical process where the model uses its capabilities to create its own "agentic frameworks" (agent harnesses).
In practice, this means that during the development of M2.7, the model itself created dozens of complex skills to assist in reinforcement learning experiments. The model thus modifies its own memory and optimizes its learning processes based on the results of its own experiments. This closed learning cycle leads to the model constantly improving without the need for massive manual code rewriting by human engineers.
What are agentic workflows?
For laypeople, the term agentic workflows can be confusing. Imagine the difference between a calculator and an accountant. A calculator (traditional LLM) will answer a question if you ask it. An accountant (AI agent), however, will not limit themselves to your question "How much did I spend on energy?". They will look at invoices, perform calculations, compare them with the budget, and then prepare a report for you. M2.7 is designed precisely for this kind of autonomous action, where it uses Agent Teams, dynamic tool searching, and complex memory.
Benchmarks: How does M2.7 compare to market leaders?
To understand the true performance of M2.7, we need to look at hard data. In the field of software engineering, which is one of the most demanding tasks for AI, the model achieved excellent results. In the SWE-Pro benchmark, it achieved a score of 56.22%, which is very close to the top results of Anthropic's Claude 3 Opus model.
Here is a brief comparison in key areas:
- Software Engineering (SWE-Pro): M2.7 (56.22%) vs. top models like Claude Opus (very close level).
- Office Tool Usage (GDPval-AA): M2.7 achieves a score of 1495, which is the highest value among all open-source models.
- Complex Engineering Systems (Terminal Bench 2): 57.0%.
- Skill adherence: 97% when working with more than 40 complex skills.
These results indicate that M2.7 is not just a "fun chat application," but a robust tool for professionals who need to integrate AI into real-world workflows, such as log analysis, bug troubleshooting, or automation in Excel and Word.
Practical Impact: What does this mean for businesses and users?
The integration of MiniMax M2.7 on platforms from NVIDIA is crucial for scalability. Thanks to optimization for NVIDIA hardware, companies can deploy these complex agentic systems at scale without performance degradation as the number of tasks increases.
Usage possibilities:
- Software Development: From analyzing entire projects to securing code. AI can function as an autonomous junior programmer.
- Administration and Office: M2.7 handles complex edits in Excel, PPT, and Word, including multi-round revisions. This can significantly reduce time spent on routine document work.
- Creative and Emotional Interactions: The model exhibits a high degree of character consistency and emotional intelligence, opening the door for more advanced virtual assistants or game characters.
Availability, Price, and Situation in the Czech Republic
In terms of availability, MiniMax M2.7 is primarily accessible via API and cloud platforms integrated into NVIDIA infrastructure. Currently, no specific price for the end Czech user is listed, but the standard model for these technologies is based on "pay-as-you-go" (payment per token/usage). For Czech companies, it is crucial to monitor whether the API provider guarantees compliance with the EU AI Act, which is essential in Europe for the legal deployment of autonomous systems.
Availability in Czech: While the basic capabilities of MiniMax large models usually include Czech, for complex agentic tasks (such as programming or deep document analysis), English is the primary language. However, for the Czech market, it is important that the model's ability to understand context and work with data is already at a level that allows even Czech-speaking developers to use its assistance when working with English code or documentation.
Is MiniMax M2.7 safe for use in European companies given the regulations?
That depends on how the model is deployed. Because M2.7 has autonomous capabilities (agentic workflows), it falls under stricter categories within the EU AI Act. Companies must ensure that agent-driven processes are transparent and under human supervision to meet European standards for high-risk AI.
Can M2.7 replace a human programmer?
Not directly. M2.7 is designed as an extremely powerful assistant. It can independently solve errors, analyze logs, and deliver entire projects, but it still requires a human architect to define goals, control security, and perform final validation of results.
How does M2.7 differ from a regular ChatGPT?
While ChatGPT is primarily a chat interface focused on conversation, M2.7 is optimized for "agentic" behavior. This means it focuses more on actually *completing* a task (e.g., fixing code and testing it), rather than just telling you how you should do it.