Hugging Face published an announcement on October 8, 2026, detailing six custom AI models created using an agent named ML-intern. The projects involved building domain-specific models, character adapters, camera-angle additions, and smaller student models designed to run on resource-constrained hardware.
The first project produced a citrus disease visual language model. Using 3,017 annotated images merged from three Project-AgML sources, the model fine-tuned Qwen3.5-2B to identify 21 pests, illnesses, and nutritional gaps. The base model achieved a 14.9 percent score on 335 test photos, while the fine-tuned version reached 52.8 percent after two epochs on one A10G. Compute cost totaled about USD 1.90.
The second project created a LoRA on FLUX.2 klein base 4B to draw the Huggy character using 84 captioned drawings from the Chunte/huggy_for_training dataset. The agent saved a checkpoint every 100 steps. Step 200 produced the first fully on-model output, while step 500 introduced style bleeding. Compute cost reached about USD 7.60.
Camera and Doodle Tools
The third project added camera-angle capabilities by rendering 1,030 household objects from Google Scanned Objects at 24 angles each. Training ran for 2,000 steps on an A100. Total compute cost for the viewpoint orbit project came to about USD 16. A separate doodle-in LoRA used 6,042 training pairs built from Open Images, substituting objects removed by a LaMa inpainting model with a magenta scribble. Training ran for 2,000 steps on one A100 at a compute cost of about USD 24.
The fourth project built a small prompt rewriter. A 9B teacher model rewrote 8,797 image requests, from which 1,840 examples were selected. Training 0.8B and 2B student models took 12 and 18 minutes on an A10G for USD 0.75, with the 0.8B version also shipped as an 812 MB GGUF file. A second distillation project took Logolabs' Agate Preview 002 and reduced its step count from 50 steps down to four steps across two training runs totaling about USD 37 in compute costs.
GitHub repository yvrjsharma/ml-intern-prompts hosts all seven prompts used in these projects.
