# Hugging Face releases ML-Intern, its open-source agent for the model-training loop > Source: > Published: 2026-06-19 02:52:58+00:00 [Hugging Face](https://huggingface.co/?ref=runtimewire)s Lewis Tunstall and Aksel Joonas Reedi moved [ML-Intern](https://github.com/huggingface/ml-intern/tree/main?ref=runtimewire) out of beta on June 18, releasing an open-source agent that tries to automate the machine learning research loop from literature review to training runs. [Lewis Tunstall on X](https://x.com/_lewtun/status/2067614409678020999?ref=runtimewire) [Aksel Joonas Reedi on X](https://x.com/akseljoonas/status/2046543093856412100?ref=runtimewire) The release matters because it is not framed as another general-purpose coding assistant. In a post on X, Tunstall wrote that "ML Intern is out of beta!" and said the tool had been used more than 12,000 times since launch to create hundreds of models and datasets on the Hugging Face Hub. He also said those runs generated more than 300 million tokens, mostly through Kimi K2.6. Those are company-side usage claims, not independently audited metrics, but they explain why Hugging Face is moving the project from experiment to public infrastructure. Reedi, whose name appears in the repositorys citation block as one of the projects authors, described ML-Intern as "an open-source implementation of the real research loop" used by Hugging Face researchers. The repositorys suggested citation lists several Hugging Face contributors as authors, making this a Hugging Face research-infrastructure release rather than a standalone startup launch. ### The Hugging Face playbook, applied to agents Hugging Face has built its position by turning internal machine learning plumbing into shared public infrastructure. ML-Intern follows that pattern. The project is less about inventing a new destination than making the existing Hugging Face surface area - models, datasets, papers, Spaces, Jobs, inference providers and open-source libraries - easier for a single agent to operate. That is the strategic point. A coding agent can write a fine-tuning script. ML-Intern is designed to decide what should be trained, find the supporting papers, inspect datasets, reformat them, run jobs, read evaluation outputs and try again. The project is pitched as an ML intern that autonomously researches, writes, and ships ML-related code across the Hugging Face ecosystem, with access to docs, papers, datasets and cloud compute. The tool is available in two forms: a command-line interface from the [GitHub repository](https://github.com/huggingface/ml-intern/tree/main?ref=runtimewire) and a hosted [web app on Hugging Face Spaces](https://huggingface.co/spaces/smolagents/ml-intern?ref=runtimewire). The CLI runs locally, and the Space offers a web and mobile UI. That last detail is the business subtext. ML-Intern is open source, but the workflow it encourages can consume Inference, Spaces, Jobs and GPUs. Hugging Faces homepage says Team and Enterprise plans start at $20 per user per month, Inference Providers expose access to more than 45,000 models through a unified API, and GPU compute starts at $0.60 per hour. ML-Intern is a developer tool, but it is also a demand generator for the paid compute layer sitting beneath the Hub. ### What ML-Intern actually does Per Reedis launch post, a user gives ML-Intern a prompt and the agent researches papers, follows citations, implements ideas in GPU sandboxes, iterates and builds models for a given use case. He said the agent can search arXiv and [Hugging Face Papers](https://hf.co/papers?ref=runtimewire), pull datasets from [Hugging Face Datasets](https://hf.co/datasets?ref=runtimewire), inspect and reformat data before training, launch training jobs when local GPUs are unavailable via Hugging Face Jobs, monitor runs, read evaluation outputs and retrain. The release also shows Hugging Face leaning into agents as a layer on top of its existing open-source stack. RuntimeWire [reported earlier this month](/article/hugging-face-build-small-hackathon-small-models) that Hugging Face was pushing developers [toward small-model efficiency](/article/hugging-face-build-small-hackathon-small-models) through its Build Small Hackathon track. ML-Intern is adjacent to that same thesis: if smaller and open-weight models are becoming useful enough to run locally or cheaply, the scarce resource shifts from model access to workflow orchestration, evaluation discipline and dataset handling. ### The benchmark claims are useful, but still claims Reedi gave three examples of ML-Interns behavior in the launch post. In scientific reasoning, he said the agent found OpenScience and NemoTron-CrossThink, added seven difficulty-filtered dataset variants from ARC, SciQ and MMLU, and ran 12 supervised fine-tuning jobs on Qwen3-1.7B. He claimed that moved a GPQA score from 10% to 32% in under 10 hours, compared with a 22.99% result he attributed to Claude Code. In healthcare, he said ML-Intern inspected available datasets, judged them too low quality, wrote a script to generate 1,100 synthetic data points for emergency, hedging and multilingual cases, then upsampled the data 50x for training. He claimed that run beat Codex on HealthBench by 60%. In competitive math, he said the agent wrote a GRPO script, launched A100 training on [Hugging Face Spaces](https://hf.co/spaces?ref=runtimewire), observed reward collapse and ran ablations until it found a working path. Those examples are the right way to understand Hugging Faces ambition, but they should not be read as independently validated leaderboards. The claims come from the launch post. The stronger verified fact is narrower and more important: Hugging Face has shipped an open-source agent with direct hooks into the places machine learning work already happens on its platform. That gives ML-Intern a different wedge from coding assistants like Claude Code or Codex. Those tools are primarily positioned as coding assistants. Hugging Face is positioning ML-Intern as a post-training assistant that knows where model builders live: papers, citations, datasets, notebooks, training jobs, evaluation logs and model repos. The product bet is that ML research automation is not one task. It is a chain of small, failure-prone tasks that become valuable when an agent can keep context across them. ### Open source as distribution, compute as the meter Hugging Face says its Hub hosts more than 2 million models, more than 1 million applications and more than 500,000 datasets, with more than 50,000 organizations using the platform. Its public open-source stack includes Transformers, Diffusers, Safetensors, Datasets, TRL, PEFT, Accelerate and smolagents. ML-Intern is another open-source artifact, but it is also a routing layer into that inventory. The unanswered question is not whether ML researchers want agents to remove busywork. They do. The question is where users will trust autonomy. Letting an agent read papers and draft training scripts is low risk. Letting it spend GPU budget, transform datasets, generate synthetic healthcare data or choose benchmark methodology requires more review. The emphasis on running in sandboxes, launching jobs when local GPUs are unavailable, and monitoring runs shows Hugging Face is aware of that boundary. The release is still one day old, but the direction is clear. Hugging Face is taking the open-source community infrastructure it built for sharing models and turning it into an execution environment for agents. If ML-Intern works as advertised, the intern is not replacing the researcher. It is moving the researcher closer to being a manager of experiments, compute budgets and evaluation criteria - exactly the layer Hugging Face already wants to own.