Hugging Face AI Updates: October 9, 2026

1. Hugging Face Releases Six Models Built End to End by ML Intern for About $103

Hugging Face. Hugging Face published a case study in which its ML Intern agent, running in HuggingChat, planned, budgeted, smoke-tested, trained, evaluated and published six models from plain-language prompts for roughly $103 in total compute, with the results released under the ML-Intern-lab organization. The set includes Citrus Doctor, a Qwen3.5-2B fine-tune on 3,017 citrus images that raised test accuracy from 14.9% to 52.8% for about $1.90; two Qwen-Image 2.1 LoRAs for camera orbiting and scribble-to-object editing; a FLUX.2 klein style LoRA; and Pocket Rewriter, 0.8B and 2B prompt rewriters distilled from a 9B teacher, with the 0.8B model shipping as an 812 MB GGUF for CPU use. A sixth run distilled the 260M Logolabs Agate image model from 50 to 4 sampling steps and exported it to ONNX for the browser, reaching a GenEval score of 0.536 against the teacher’s 0.563. Source