Moondream has launched Lens, a fine-tuning service designed to enhance the accuracy of visual language models (VLMs) as they transition from lab settings to real-world applications. Announced on April 21, Lens is a pay-per-use API that supports both reinforcement learning and supervised fine-tuning, achieving significant improvements with minimal data. For instance, in NBA live broadcasts, fine-tuning improved the F1 score from 28% to 79%, reducing false positives significantly. The service also demonstrated superior performance in identifying countries from street view images and in medical imaging tasks, surpassing existing models like GPT-5.4. Moondream's first partner, PTZOptics, will integrate Lens for enhanced target tracking and anomaly detection. This launch complements Moondream's Photon inference engine, addressing both accuracy and speed challenges in VLM deployment.