---
base_model:
- Qwen/Qwen3-32B
datasets:
- open-thoughts/OpenThoughts-Agent-SFT-100K
library_name: transformers
license: apache-2.0
model-index:
- name: OpenThinkerAgent-32B
  results: []
pipeline_tag: text-generation
tags:
- agents
- terminal
- code
- software-engineering
---
<p align="center">
    <img src="https://huggingface.co/datasets/open-thoughts/OpenThoughts1-Agent-SFT/resolve/main/ota-logo.png" width="50%">
</p>

<p align="center">
<a href="https://www.openthoughts.ai/blog/agent" style="margin-right: 24px;">Project</a> |
<a href="https://github.com/open-thoughts/OpenThoughts-Agent" style="margin-right: 24px; margin-left: 24px;">Code</a> |
<a href="https://huggingface.co/collections/open-thoughts/openthinker-agent" style="margin-left: 24px;">Collection</a>
</p>


# OpenThinkerAgent-32B

**OpenThoughts-Agent** is an open-source effort to curate the best datasets for training agents. Our release includes [datasets](https://huggingface.co/collections/open-thoughts/openthinker-agent), [models](https://huggingface.co/collections/open-thoughts/openthinker-agent) and our [research codebase](https://github.com/open-thoughts/OpenThoughts-Agent).

[OpenThinkerAgent-32B](https://huggingface.co/open-thoughts/OpenThinkerAgent-32B) is post-trained from [Qwen/Qwen3-32B](https://huggingface.co/Qwen/Qwen3-32B) with full-parameter SFT on the **100,000-example** [OpenThoughts-Agent-SFT-100K](https://huggingface.co/datasets/open-thoughts/OpenThoughts-Agent-SFT-100K) dataset (Top-4 task sources, GLM-4.7-AWQ teacher in the terminus-2 harness, ≥5-turn trace filter). It is the flagship **OpenThinkerAgent-32B**, the strongest open-data 32B model on the average of seven agentic benchmarks.

- **Homepage:** https://www.openthoughts.ai/blog/agent
- **Repository:** https://github.com/open-thoughts/OpenThoughts-Agent

# Performance

Evaluated in the **terminus-2** harness (pass@1, mean over 3 stochastic re-runs):

| Model | Harness | SWE-Bench-Verified-100 | OpenThoughts-TBLite | Terminal-Bench 2.0 |
| --- | --- | --- | --- | --- |
| [Qwen/Qwen3-32B](https://huggingface.co/Qwen/Qwen3-32B) | Terminus-2 | 26.7 | 13.7 | 7.5 |
| **OpenThinkerAgent-32B** | Terminus-2 | 55.7 | 41.3 | 26.2 |

Across the full seven-benchmark suite (best harness per benchmark), **OpenThinkerAgent-32B** is the strongest open-data model at the 32B scale:

| Benchmark | Accuracy |
| --- | --- |
| SWE-Bench-Verified | 54.0 |
| Terminal-Bench 2.0 | 26.2 |
| Aider-Polyglot | 32.4 |
| BFCL-Parity | 85.9 |
| MedAgentBench | 47.8 |
| GAIA-127 | 23.6 |
| FinanceAgent-Terminal | 44.0 |
| **Average (7)** | **44.8** |

# Data

The model is trained on [OpenThoughts-Agent-SFT-100K](https://huggingface.co/datasets/open-thoughts/OpenThoughts-Agent-SFT-100K): (task, agent-trajectory) pairs from the Top-4 task sources (SWE-Smith, StackExchange-SuperUser, StackExchange-Tezos with synthetic augmentation, IssueTasks). Trajectories are generated by GLM-4.7-AWQ in the terminus-2 harness and filtered to traces with at least 5 model turns.

### Training hyperparameters
- learning_rate: 4e-05
- lr_scheduler_type: cosine, warmup_ratio 0.1
- global_batch_size: 96
- num_epochs: 5
- cutoff_len: 32768
- precision: bf16, DeepSpeed ZeRO-3

# Links
- 🌐 [OpenThoughts-Agent project page](https://www.openthoughts.ai/blog/agent)
- 💻 [OpenThoughts-Agent GitHub repository](https://github.com/open-thoughts/OpenThoughts-Agent)
- 📚 [OpenThinker-Agent collection](https://huggingface.co/collections/open-thoughts/openthinker-agent)
- 🧠 [Training dataset: OpenThoughts-Agent-SFT-100K](https://huggingface.co/datasets/open-thoughts/OpenThoughts-Agent-SFT-100K)

# Citation
```
@misc{openthoughts-agent,
  author = {Team, OpenThoughts-Agent},
  title = {{OpenThoughts-Agent: Data Recipes for Agentic Models}},
  howpublished = {https://www.openthoughts.ai/blog/agent},
  year = {2026}
}
```