Hugging Face reports low-cost custom models built with ML Intern, with accuracy trade-offs

Hugging Face reports using ML Intern to build six custom-model projects for about USD 103 in CPU and GPU job charges, suggesting a lower compute-cost route to specialized models rather than a measure of total development cost. Source

Hugging Face says its citrus model improved from 14.9% to 52.8% accuracy on 335 test photos, while a 0.8B prompt rewriter runs on a CPU. These reported results suggest specialization and smaller deployment footprints can be practical, but the citrus score still leaves substantial errors. Source

Hugging Face also reports failed jobs requiring resubmission and character-style spillover with longer LoRA training, underscoring the need for baseline tests and checkpoint selection. Whether these vendor-reported costs and results generalize to other datasets, hardware and real-world workloads remains unverified; they are not independent confirmation. Source