Perceptron Mk1

Perceptron Mk1 offers advanced video understanding and embodied reasoning through production APIs. Key capabilities include: • Temporal grounding with video clips and structured timecodes • In-context learning via multimodal prompting • Dynamic frame rate analysis across a 32K token context • Advanced image reasoning with pointing and counting • Focus on physical-world tasks at high volume This model is engineered for continuous event streams, providing detailed breakdowns of actions and events over time, such as in sports broadcasts or cooking videos. It can identify specific moments, auto-clip events like grasp attempts, and flag occurrences like restocking in warehouses, all from a single prompt. Its hybrid reasoning allows users to toggle reasoning off when not required. Mk1 excels at in-context learning, enabling users to present a reference image or video for it to identify matching instances across new media without the need for fine-tuning or labeled datasets. It supports side-by-side comparisons of media, streamlining workflows for defect detection, asset tracking, and visual search. Additionally, it enhances image reasoning with precise pointing and accurate counting in complex scenes, meeting the demands of production vision pipelines. Ideal for engineering teams, developers, and businesses operating in robotics, surveillance, manufacturing, and logistics, particularly those requiring robust video analysis for monitoring, quality control, and operational optimization.