Simple terminal-using agents.
💻 Code · 🤗 Models & Data · 📜 Paper · 📓 Blog
Tmax is our project around training simple, powerful terminal using agents. This codebase covers data generation, training, and evaluation. We use this to train our 'tmax' series of models, with strong performance as shown below! Please refer to our paper for more details.
Below, we give a quick overview of the codebase and how to use it.
- [3/11/26] TMax v1.1 released! We have updated the codebase to improve stability and our overall recipe. Find the models and data in our Hugging Face collection and the tasks on Harbor Hub. Please consult PR #11 for the full set of changes. To reproduce TMax v1.0 training, please use commit
6d3d606. - Initial release of the codebase and models! Please read our paper for more details.
The repo is organised around four stages of building a terminal agent:
| Stage | Where | What it does |
|---|---|---|
| Data generation | rl_data/ |
A simple, scalable, diverse, difficulty-aware pipeline for synthesising terminal-agent tasks, solving them at pass@k, analysing the corpus, and publishing it to the Hugging Face Hub. Tasks are sampled as an independent product of structured axes and packaged as self-contained Apptainer/Docker environments with programmatic verifiers. |
| Agent | Vanillux2Agent/ |
The Harbor agent used for solving and evaluation: a direct LiteLLM agent built on the vanillux prompt harness (mini-SWE-agent-derived prompts, bash tool schema, submit marker, format-error recovery, and output truncation), executing commands through Harbor's active environment. |
| Training | training/open-instruct/ |
A fork of open-instruct with fixes for Qwen 3.5 and terminal-agent training. SFT and DPPO RL launch scripts for the tmax models live under training/open-instruct/scripts/tmax/. |
| Evaluation | scripts/ + beaker_configs/ |
Shell/Slurm launchers and a Beaker pipeline that serves a model with vLLM and runs Harbor datasets (Terminal-Bench, TB-Lite, SWE-bench) against it. |
Python is run via uv; all commands run from
the repo root.
# Install dependencies
uv sync# 1. generate a small task corpus
NUM_TASKS=10 OUT_DIR=rl_data/output/tasks_smoke \
bash rl_data/scripts/generate_tasks/run_generate_tasks.sh
# 2. solve the tasks with an LLM agent at pass@k
TASKS_DIR=rl_data/output/tasks_smoke \
bash rl_data/scripts/generate_solutions/run_generate_solutions.sh
# 3. analyse pass@k + composition/balance stats
TASKS_DIR=rl_data/output/tasks_smoke \
bash rl_data/scripts/analyze/run_analyze.shrl_data/README.md is the entry point for the data side
of the project: it walks through the compositional sampler (the orthogonal
axes that define a task), the four-stage pipeline
(generate_tasks → generate_solutions → analyze → upload_to_hf), the corpus
kinds (legacy / sft_v2 / rl_v2), the SFT warm-start set, and how to
evaluate on our published Harbor dataset. Start there for anything data-related.
SFT and DPPO RL are run via the open-instruct fork in training/open-instruct/.
Launch scripts for the tmax models live under training/open-instruct/scripts/tmax/
(SFT/ and RL/):
# from training/open-instruct/, with BEAKER_CLUSTER configured, e.g. Qwen3.5-4B
bash scripts/tmax/RL/qwen35_4b.sh <beaker-image>See training/open-instruct/scripts/tmax/README.md
for runtime configuration, Sandfleet execution, dataset overrides, and pretokenized SFT input.
Run a Harbor dataset against a locally served model on Beaker, with
BEAKER_CLUSTER configured for your environment:
./beaker_configs/launch_eval.sh TMaxxx/tmax-9b-v1.1 \
--revision main \
--name tmax-v1.1-9b \
--tool-call-parser qwen3_xml \
--cluster "$BEAKER_CLUSTER" \
--dataset [email protected]See scripts/beaker/README.md for the full eval
pipeline, flags, and troubleshooting.
To run an eval yourself on a node with GPUs (no Beaker), serve the model with vLLM and point Harbor at it:
# 1. serve the model on localhost:8008 (a separate shell/process)
uvx vllm==0.19.1 serve TMaxxx/tmax-9b-v1.1 \
--served-model-name tmax-v1.1-9b \
--enable-auto-tool-choice --tool-call-parser qwen3_xml \
--tensor-parallel-size 8 --port 8008
# 2. set the Daytona API key
export DAYTONA_API_KEY='xxx'
# 3. run a Harbor dataset against the local vLLM endpoint
uv run harbor run \
--dataset [email protected] \
--env daytona \
--agent-import-path Vanillux2Agent:Vanillux2Agent \
--model openai/tmax-v1.1-9b \
--agent-kwarg api_base=http://localhost:8008/v1 \
--n-concurrent 16 \
-k 5 \
--job-name tmax-v1.1-9b-terminal-bench-2.0Feel free to swap out daytona for your sandboxing backend of choice.
The 13,565 hardened tasks used by TMax v1.1 are public on Harbor Hub as
tmax/tmax-13k-v1.1.
The v1.1 and latest tags resolve to the same complete release. Its pinned
task references preserve the instructions, verifiers, supporting test assets,
and Docker wrappers from the
Hugging Face Harbor export.
The training data and
metadata companion
cover the same task IDs.
# Evaluate an agent on a 10-task subset using Docker.
# Set MODEL to your provider/model and configure its API credentials.
uvx --from harbor==0.25.0 harbor run -d "tmax/[email protected]" \
--agent terminus-2 --model "$MODEL" --env docker -l 10Drop -l to run the full set, or use -i/-x to include/exclude tasks by name.
Per-task rewards and agent/verifier logs land under jobs/<job-name>/. See
rl_data/README.md
for the full set of run options.
The earlier 14,601-task release remains available as
tmax/TMax-15K-Harbor.
It contains the original tasks before the v1.1 audit, repairs, and exclusions.
uvfor dependency management (deps pinned inpyproject.toml/uv.lock).- An LLM API key for the configured model (e.g.
GEMINI_API_KEY); local vLLM / Ollama / OpenAI-compatible endpoints are also supported via env vars. - (for data generation)
apptaineron PATH for building and running task containers. - (for training) A configured Docker/Podman, Apptainer, or Sandfleet backend. Docker Hub credentials may be needed for authenticated or high-volume image pulls; Apptainer and Sandfleet can use prepared SIF images. See the training recipes for setup.
HF_TOKENfor the Hugging Face upload stage and for pulling gated models.- (to evaluate on the published Harbor datasets) a model API key
for your chosen agent plus a container runtime — Docker locally, or
DAYTONA_API_KEYfor the Daytona sandbox on Docker-less hosts.
- V1.1:
- Models and data: TMax v1.1 collection.
- Harbor dataset:
tmax/[email protected](13,565 tasks). - Evaluation rollouts:
evaluation_assets/v1.1/(2B, 4B, 8B, 9B, and 27B on Terminal-Bench 2.0 using Daytona).
- V1.0:
- Models and data: Ai2 TMax collection.
- Evaluation rollouts:
evaluation_assets/v1.0/. - Training code:
6d3d606.
This codebase is licensed under Apache 2.0 as given in LICENSE.
The TMax v1.1 training data, metadata, and Harbor export use ODC-BY, following the original TMax dataset. Research and educational use should follow Ai2's Responsible Use Guidelines. Gemini 3.1 Pro outputs included in the data remain subject to Google's Terms of Service.
If you use this codebase or models in your research, 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},
}