Perceptron Mk1.5: Multimodal and Embodied Agent Guide

Perceptron Mk1.5 was announced on September 25, 2026 for embodied agents and visual reasoning. Its feature set is relevant to applications that need to reason about locations or events across images, video, and audio, then produce an action or structured annotation.

OpenRouter lists 36,864 context tokens for this route, while Perceptron’s announcement describes 32K tokens of multimodal context. Those are distinct published limits. Use the limit of the actual endpoint, and reserve space for every input modality and the output. OpenRouter route · Perceptron’s announcement

Perceptron says Mk1.5 adds native audio support, video tracking, tool use, and more complex visual reasoning. It can emit text plus points, boxes, polygons, clips, and tracks. The vendor demonstrates use on drones, quadrupeds, glasses, and phones. These are developer-reported capabilities and demonstrations; no physical-agent test was performed for this article.

For an application, validate annotation structure separately from localization accuracy. A box can have valid coordinates and still identify the wrong object. A clip can use a valid time interval and still miss the requested event. Keep reference examples whose spatial or temporal answer can be checked directly, and include cases where the appropriate result is that no target is present.

Expose actions at the level the application can inspect and validate. A model proposing a camera movement or a device action needs feedback about the resulting state before the next decision. Record the observation and action together so a failure can be traced to perception, selection, or execution.

A broad creative-coding prompt would test a different capability than tracking a subject through video. The model could produce a pleasant browser demo without establishing that its embodied reasoning works on a deployed device. An evaluation should follow the application’s actual observation and action loop.

This route was paid when checked, so it did not receive the SVG, globe, and wizard-game prompts in the free-model run. At the base rates, a hypothetical 10,000 input tokens and 1,000 output tokens cost $0.003 before extra operations. That is a pricing illustration, not a measured sensor or robotics workflow.