💻 Code · 🤗 Models & Data · 📜 Paper · 📓 Blog

TMax v1.1 27B

This terminal-agent model was trained using DPPO on Qwen/Qwen3.6-27B with the TMax-13K v1.1 task set.

This model is part of the TMax v1.1 collection. The original TMax recipe is described in our paper.

Checkpoints

This branch contains step 200. The main branch contains step 200, the highest-scoring available checkpoint in the completed TB-Lite sweep. Checkpoint selection uses TB-Lite, not Terminal-Bench 2.1.

Checkpoints 100, 200, 300, 400, 500 are provided as branches named step-0100, step-0200, etc.

Evaluation Results

Training step	TB-Lite (%) ± SE (pp)	Selection
100	75.08 ± 1.81
200	77.35 ± 1.78	main
300	73.85 ± 1.67
400	74.66 ± 1.81
500	74.83 ± 1.51

The selected checkpoint scores 43.94 ± 1.56% on Terminal-Bench 2.1 (88 tasks, three trials per task; configure-git-webserver excluded).

On Terminal-Bench 2.0 it scores 44.94 ± 1.76% (all 89 tasks, three trials per task, Daytona backend).

These scores use Vanillux 0.2.3 with preserved intermediate thinking and the TB-Lite checkpoint sweep completed in October 2026. The shipped chat_template.jinja exactly matches the template used for these evaluations: thinking is enabled by default, and previous reasoning is retained across tool responses and user reprompts. No chat-template override is required. An explicit enable_thinking=false can disable thinking. TB-Lite and Terminal-Bench 2.1 use Sandfleet through Harbor; Terminal-Bench 2.0 uses Daytona. All scores report three trials per task and standard errors in percentage points. release_manifest.json records the preserved-thinking sweep, and preserved_thinking.json records the exact template hash.

Model Details

Model Description

Developed by: Ai2
Language: English
License: Apache 2.0
Finetuned from: Qwen/Qwen3.6-27B
Task set: TMaxxx/tmax-13k-v1.1
Training method: DPPO reinforcement learning

Use

Serve the model with vLLM and use the terminal-agent harness in our codebase:

Use a Linux NVIDIA GPU environment with a CUDA toolkit matching PyTorch (CUDA_HOME set and nvcc on PATH; see the vLLM installation guide). The tested environment uses Python 3.12, PyTorch 2.13.0, and CUDA 13.0.

uvx --python 3.12 vllm==0.32.0 serve TMaxxx/tmax-27b-v1.1 \
  --served-model-name tmax-v1.1 \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_xml \
  --port 8008 \
  --max-model-len 65536 \
  --tensor-parallel-size 1

The example uses one GPU. Increase --tensor-parallel-size or reduce --max-model-len to fit your available GPU memory.

The Qwen3.5/3.6 checkpoints contain the text-only causal language model; use the Qwen3 XML tool parser.

Training Details

Released training step: 200
Main checkpoint: 200, selected on TB-Lite
Available milestone checkpoints: 100, 200, 300, 400, 500

The source checkpoint is TMaxxx/tmax-27b-v1.1 at step 200. Model weights and training provenance are preserved, including campaign_provenance.json where supplied. See release_manifest.json for the immutable source commits and evaluation results. See the training launch scripts and original provenance for run-specific settings.

License

This model is licensed under Apache 2.0. It is intended for research and educational use in accordance with Ai2's Responsible Use Guidelines.

Citation

If you use our model or data, please cite our paper:

@misc{ivison2026tmaxsimplerecipeterminal,
      title={Tmax: A simple recipe for terminal agents}, 
      author={Hamish Ivison and Junjie Oscar Yin and Rulin Shao and Teng Xiao and Nathan Lambert and Hannaneh Hajishirzi},
      year={2026},
      eprint={2606.23321},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2606.23321}, 
}