Researchers have introduced TMax—an open method for training agentic models to operate in the terminal using Reinforcement Learning (RL). The approach is based on creating a database of 15,000 complex tasks (TMAX-15K), which allows abstract skills to be transformed into measurable goals with high-quality feedback.

image
image

What Happened

As part of the research, a model with 9 billion parameters was trained, demonstrating a 27% result on the TerminalBench-2.0 benchmark. The TMax method uses Reinforcement Learning to master complex terminal operations, which is more efficient than the standard pre-training stage.

Context

Traditional AI development often focuses on increasing the number of parameters and data volumes. The TMax method proposes a paradigm shift: moving from scaling weights to qualitatively training specific skills (reasoning, verification, tool-use) through structured tasks.

Why It Matters for the Industry

This technology paves the way for creating efficient Small Language Models (SLMs) capable of competing with large proprietary systems in narrow agentic scenarios. This allows companies to prototype specialized terminal agents faster without requiring massive computational resources.

Why It Matters for Users

For developers and specialists, this means the possibility of creating more controllable and predictable AI agents. Development is shifting toward high-quality training "recipes," making the creation of effective assistants for DevOps and system administration more accessible.

What Is Not Yet Known / Limitations

The current result of 27% in TerminalBench-2.0 indicates a significant technological gap between the research prototype and a production-ready solution.

Sources

Author

Look at AI, Editorial Staff