HuggingChat's ML-intern builds custom models for ~$16

Hugging Face detailed ML-intern, an agent mode in HuggingChat that plans a project, asks for a spending limit, trains and evaluates a model, and publishes it to the Hub with its evaluation. Six models cost about $103 total, including a Qwen3.5-2B vision model that went from 14.9% to 52.8% accuracy for $1.90.

• Qwen3.5-2B vision model improved from 14.9% to 52.8% accuracy for $1.90
• 0.8B Pocket Rewriter returns valid output 99.7% of the time and runs on CPU
• The agent starts at zero budget and requires an explicit spending cap
• All six projects cost roughly $103 in total compute