ml-intern Open-Source ML Coding Agent
Scientific reasoning (GPQA, Qwen3-1.7B): 10% to 32% in under 10 hours, outperforming Claude Code’s best of 22.99% on the same task. For competitive math, it wrote a full GRPO script, watched the reward collapse mid-run, and ran ablations until it recovered — no human intervention.
Within hours of launch, one user asked it to apply LoopLM to BitNet 1.58 and it implemented and began training from a single sentence.
What the community is flagging: results are self-reported with no independent replication at launch. PostTrainBench has documented cases of agents gaming their own eval — none confirmed for ml-intern specifically. One researcher called it a “hyper-hyper param sweep” — strong at combining existing methods, unproven on genuinely novel techniques.
CLI: github.com/huggingface/ml-intern Web: huggingface.co/spaces/smolagents/ml-intern Community models: huggingface.co/ml-agent-explorers"
TL;DR
ml-intern is an open-source AI coding agent purpose-built for machine learning research and training workflows. It reads live Hugging Face documentation at runtime before generating code, reducing hallucination of deprecated API parameters in libraries like TRL. General-purpose coding agents frequently hallucinate outdated API calls when working with fast-moving ML libraries. ml-intern solves this by grounding its code generation in live HF documentation, and it operates autonomously — writing, executing, observing, and ablating training runs without human intervention.
What it actually is
- What: ml-intern is an open-source AI coding agent purpose-built for machine learning research and training workflows. It reads live Hugging Face documentation at runtime before generating code, reducing hallucination of deprecated API parameters in libraries like TRL.
- Who built it / maintained by: Hugging Face (with contributions from the ml-agent-explorers community; post attributed to Arnav Jaitly / arnitly)
- Status: beta
- Why it matters: General-purpose coding agents frequently hallucinate outdated API calls when working with fast-moving ML libraries. ml-intern solves this by grounding its code generation in live HF documentation, and it operates autonomously — writing, executing, observing, and ablating training runs without human intervention.
- How it compares to alternatives:
- Claude Code
- GitHub Copilot
- Cursor
- OpenDevin
- SWE-agent
- Aider
- GitHub stars: 8,483 · License: unknown · Archived: no
Links
- Repo: https://github.com/huggingface/ml-intern
- Docs: https://huggingface.co/spaces/smolagents/ml-intern
Kickstarter guide
Clone the repository from github.com/huggingface/ml-intern and follow the CLI setup instructions. You can also try the tool immediately via the hosted web UI at huggingface.co/spaces/smolagents/ml-intern with no local install required. To run a training experiment, provide a single natural-language instruction (e.g., 'fine-tune Qwen3-1.7B with GRPO on math reasoning') and the agent will fetch current HF docs, write the script, and iterate autonomously. Community-contributed models and experiment checkpoints are available at huggingface.co/ml-agent-explorers for reference baselines.