How finetuning works

Detect which player is currently holding the basketball in NBA game footage. The base model fires dozens of false-positive boxes. After reinforcement learning fine-tuning, the model locks onto the correct player with an F1 of 0.79, compared to 0.53 from GPT-5.4.

Detect State Farm logos in NBA broadcast footage. The base model produces 30 false positives and misses 18 logos across the test set. After fine-tuning, F1 reaches 1.0 with zero false positives and zero false negatives.

Predict the country from a single street-view image by reading road markings, signage, and vegetation. SFT fine-tuning on a small dataset takes Moondream from 28.6% to 71.1% accuracy across 53 countries, outperforming GPT-5.4 at 69.8%.

Classify real photos of hand gestures as rock, paper, or scissors. With only 5 training examples per class and 50 RL steps, accuracy jumps from 54.8% to 98.8%. This demo highlights extreme data efficiency: a useful model from almost no training data.

Classify retinal fundus images into three glaucoma stages: normal, early, or advanced. The base model achieves only 17.6% accuracy. After 100 RL steps, the fine-tune reaches 69.2%, more than double GPT-5.4's 33.2%.

Identify the action happening in a video from a 3x3 grid of frames. This teaches the model to reason across time, connecting motion across multiple frames to describe the action. SFT fine-tuning raises template match accuracy from 0% to 50%.

Localize airplanes in satellite imagery of airports using Moondream's point skill. The base model fires 1,906 false positives. After fine-tuning with tiled inference, F1 improves from 0.30 to 0.55, while GPT-5.4 achieves only 0.10.