OpenThoughts: Training Highly Capable Open-Source Agentic Models

A collaborative effort led by IFML researchers has produced an open dataset, language model, and training pipeline designed to help scientists build more capable and transparent AI agents—systems that can reason through complex tasks, use software tools, and interact with digital environments. The work, which researchers in UT Austin’s machine learning community have majorly contributed, addresses a growing challenge in artificial intelligence: while commercial companies have rapidly advanced AI agents, the data used to train them has largely remained proprietary, limiting reproducibility and open research.

The project introduces the OpenThoughts-Agent ecosystem, which includes the OpenThoughts-Agent-SFT-100K dataset, the OpenThinkerAgent-32B model, and accompanying code and evaluation resources. The work is described in a new research paper and has been publicly released through Hugging Face and the OpenThoughts initiative.

Rather than simply collecting more data, the researchers focused on identifying what kinds of training examples produce stronger AI agents. They conducted more than 100 controlled experiments comparing different data sources, filtering methods, and task types before assembling a supervised fine-tuning dataset containing 100,000 carefully curated agent trajectories. These examples teach models how to complete multi-step tasks such as software engineering, terminal interaction, and question answering that require planning, reasoning, and tool use.

The resulting OpenThinkerAgent-32B model, built by fine-tuning Qwen3-32B on the new dataset, achieved an average score of 44.8 percent across seven widely used agent benchmarks, outperforming previous open-data models of comparable size. For nonexperts, benchmark scores provide a standardized way to compare how well different AI systems perform the same collection of challenging tasks. The results suggest that the quality and diversity of training data can have a greater impact on agent performance than simply increasing the amount of data.

In announcing the release, UC Berkeley professor and Bespoke Labs co-founder Alex Dimakis wrote, “Better data is all you need,” summarizing one of the project’s central findings. The researchers also found that AI agents benefit from diverse training environments because “different domains teach different capabilities,” while some commonly used synthetic data augmentation techniques provided little measurable improvement.

By releasing the datasets, model, evaluation benchmarks, and training pipeline openly, the team aims to make AI agent research more reproducible and accessible. The researchers note that their work provides a foundation for future studies on how data quality, diversity, and curation influence the capabilities of increasingly autonomous AI agents.

Links:

https://www.openthoughts.ai/blog/agent 
https://arxiv.org/pdf/2606.24855 
https://huggingface.co/open-thoughts/OpenThinkerAgent-32B 
https://huggingface.co/datasets/open-thoughts/OpenThoughts-Agent-SFT-100K 
https://huggingface.co/collections/open-thoughts/openthinker-agent2
https://x.com/AlexGDimakis/status/2069834784168366473