Hugging Face just open-sourced ml-intern, an ML research agent that can autonomously read papers, curate datasets, kick off GPU training, evaluate results, and iterate — all without a human in the loop. Built on the company's own smolagents framework, it comes with both a CLI and a web interface, and the code is up on GitHub.
ml-intern's toolchain is tightly coupled to the Hugging Face ecosystem. It searches arXiv and HF Papers, follows citation chains to dig deeper, browses datasets on the Hugging Face Hub, checks quality, reformats data, and feeds it into training. If you don't have a local GPU, it can spin up cloud training jobs via HF Jobs, then automatically read evaluation outputs, diagnose failures, and retry. By default, it uses Claude Sonnet 4.5 to drive its decision loop, with a maximum of 300 iterations per run and automatic context compression when the token count exceeds 170k.
Hugging Face shared three case studies in its announcement. For a scientific reasoning task, the agent traced citation chains from a benchmark paper to find the OpenScience and NemoTron-CrossThink datasets, filtered seven variants from ARC, SciQ, and MMLU by difficulty, ran 12 rounds of supervised fine-tuning on Qwen3-1.7B, and boosted GPQA scores from 10% to 32% — all in under 10 hours. In a medical scenario, the agent decided existing datasets weren't good enough, wrote its own script to generate 1,100 synthetic samples, augmented them 50x, and beat Codex by 60% on HealthBench. For a competition math problem, it wrote a GRPO training script from scratch, launched training on an A100 via Hugging Face Spaces, observed reward collapse, and ran ablation experiments to diagnose the issue.