Downloads · 30 days
80
73% of all-time downloads
TMaxxx/qwen35-9b-dppo-terminal-lego
qwen35-9b-dppo-terminal-lego is a reinforcement learning model from TMaxxx. Use it for the reinforcement learning task on the model card, and read the license before you ship it in a product. The card lists the license as apache-2.0.
<p align="center" 💻 <a href="https://github.com/hamishivi/tmax"Code</a · 🤗 <a href="https://huggingface.co/collections/TMaxxx/tmax-v11"Models & Data</a · 📜 <a href="https://arxiv.org/abs/2606.23321"Paper</a ·
Downloads · 30 days
80
73% of all-time downloads
All-time downloads
110
Public
Parameters
9B
89.6 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors17.9 GB · 100%
From the Hugging Face model README

This terminal-agent model was trained using DPPO on Qwen/Qwen3.5-9B with the Terminal-Lego task set.
This model is part of the TMax v1.1 collection. The original TMax recipe is described in our paper.
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.
| Training step | TB-Lite (%) ± SE (pp) | Selection |
|---|---|---|
| 100 | 58.15 ± 2.17 | |
| 200 | 63.15 ± 2.00 | main |
| 300 | 62.41 ± 1.97 | |
| 400 | 61.64 ± 1.97 | |
| 500 | 62.04 ± 1.87 |
The selected checkpoint scores 26.52 ± 1.78% on Terminal-Bench 2.1 (88 tasks, three trials per task; configure-git-webserver excluded).
These v1.1 scores use the fixed Vanillux0.2.3 harness with Sandfleet through Harbor, three trials per task (96 TB-Lite tasks). Scores and standard errors are taken from the completed evaluation sweep updated on October 5, 2026. The harness differs from the original release; use the original model cards for v1.0 results. The per-checkpoint results and pinned source revisions are included in release_manifest.json.
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/qwen35-9b-dppo-terminal-lego \
--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.
The source checkpoint is TMaxxx/qwen35-9b-dppo-terminal-lego at step 200. Original checkpoint files 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.
This model is licensed under Apache 2.0.
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},
}
You should also cite the original paper this data came from:
@misc{yang2026terminallego,
title={What Makes Interaction Trajectories Effective for Training Terminal Agents?},
author={Sidi Yang and Chaofan Tao and Jierun Chen and Tiezheng Yu and Ruoyu Wang and Yuxin Jiang and Yiming Du and Wendong Xu and Jing Xiong and Taiqiang Wu and Lifeng Shang and Xiaohui Li and Ngai Wong and Haoli Bai},
year={2026},
eprint={2606.03461},
archivePrefix={arXiv},
primaryClass={cs.AI},
url={https://arxiv.org/abs/2606.03461}
}