allenai/tmax-27b

VISIONConcurrency Cost:2Model Size:27BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Jun 21, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

allenai/tmax-27b is a 27 billion parameter language model developed by Ai2, fine-tuned using DPPO on a Qwen 3.6 27B base for use as a terminal agent. It specializes in executing commands and interacting with terminal environments, achieving approximately 43% on Terminal Bench 2.0. This model is optimized for automated task execution within command-line interfaces, offering enhanced performance over its base model in terminal-specific benchmarks.

Loading preview...

TMax 27B: A Specialized Terminal Agent

TMax 27B, developed by Ai2, is a 27 billion parameter model specifically designed and fine-tuned for terminal agent applications. Built upon the Qwen 3.6 27B base model, it leverages Deep Proximal Policy Optimization (DPPO) to excel in command-line environments.

Key Capabilities & Performance

  • Terminal Task Execution: TMax 27B is optimized for interacting with terminal environments, making it suitable for automating system administration, development workflows, and other command-line tasks.
  • RL Fine-tuning: The model underwent 160 steps of RL training using DPPO, significantly improving its performance on terminal-specific benchmarks.
  • Benchmark Improvement: It achieves approximately 44.9% on Terminal Bench 2.1 and 42.7% on Terminal Bench 2.0, outperforming its Qwen 3.6 27B base model in these metrics.
  • Context Length: Supports a maximum overall context length of 65536 tokens, allowing for complex multi-turn interactions within the terminal.

When to Use TMax 27B

  • Automated Scripting: Ideal for scenarios requiring an AI to execute commands, navigate file systems, and perform operations within a terminal.
  • Developer Tools: Can be integrated into development environments for automated testing, deployment, or code management tasks.
  • Research in Agentic AI: A strong candidate for research into autonomous agents operating in command-line interfaces.

This model is part of a larger collection of TMax terminal agents and is licensed under Apache 2.0. For detailed evaluation methodology and training specifics, refer to the Tmax paper and the codebase.