For the EgoLongQA track of the ECCV 2026 Wearable-AI Challenge, researchers developed "Ambient," a compact vision-language model designed for long-video perception in ten-minute egocentric clips. The system achieved first place in the sub-2B parameter category with a score of 0.8279 on the held-out test set. The architecture was developed through a distillation process where the junior perception module of a larger, tool-using agentic pipeline served as the teacher. By training on teacher traces filtered specifically for correct answers, the resulting student model reached 89% of the larger pipeline's accuracy while utilizing only 1.1% of its parameters. This distillation improved the base model's performance from 27.1% to 81.4% on held-out questions. Although the initial backbone comprised 2.2132B parameters, the team brought the model under the 2B limit by pruning the multilingual embedding table from 248,320 to 143,469 rows, resulting in a final count of 1.9985B parameters. Unlike the original agentic pipeline, Ambient answers multiple-choice questions in a single greedy forward pass, providing an efficient solution for long-form video understanding.