Hugging Face's ML Intern turns a prompt into a trained model: 6 models in a week for ~$100
Gradio · x · 2026-10-12
Hugging Face's team introduces ML Intern: with it enabled in HuggingChat, a single prompt becomes a trained, evaluated public model on the Hub — the agent plans the work, asks for a compute budget before spending, smoke-tests before real training, then trains, evaluates and publishes on HF hardware.
Highlights:
- The author wanted a tiny version of Qwen-Image 2.1's prompt rewriter (official: 9B, 20GB). A day later ML Intern delivered a 0.8B CPU-runnable model: 99.7% valid output, 1/4 the teacher's tokens, total cost $16 including labeling 8,797 examples with the 9B model.
- 6 more models followed the same way in a week, at $100 total compute.
Prompting method: invest heavily in the first message (450 → 2,000 words across projects); put verified facts under a "Verified facts, do not re-derive" heading so budget goes to real work. All 7 prompts are open-sourced at yvrjsharma/ml-intern-prompts.
Related event: Hugging Face Unveils ML Intern: Prompt-to-Model Auto Training(2 posts)→
More from coding & agent
- Tejas Kumar: in the AI era, frontend value is taste and the web belongs to agents — TejasKumar_ · 2026-10-12
- Matt Pocock: Blame the codebase, not the agent, for bad tests — mattpocockuk · 2026-10-12
- Dev builds Claude-coded 'daily wearable' firmware packing 27 games and 16 utilities — ezshine · 2026-10-12
- MCP memory app keeps persistent context across Claude, ChatGPT and Codex — Jolly-Scientist-2484 · 2026-10-12
- HANDBOOK.md benchmark: top agents pass just 36.2% of 65 policy-heavy agentic tasks — bibryam · 2026-10-12
- LangChain's Harrison Chase backs the bet that AI feeds will be the form factor of 2027 — hwchase17 · 2026-10-12