Perceptron Mk1

What is Perceptron Mk1?

Perceptron Mk1 is a closed-source model from Perceptron designed for video understanding and embodied reasoning. It is intended to analyze images and video, reason across time, and produce structured outputs such as timecodes, clips, points, boxes, polygons, tracks, and text.

The model is positioned for physical AI and robotics workflows, where it can process continuous visual streams rather than isolated frames. According to the source, it matches frontier performance on image, video, and embodied reasoning tasks while being priced below some comparable frontier offerings.

Key Features

• Temporal reasoning over video: Mk1 can examine events across time and return structured breakdowns of what happened and when, which is useful for sequential tasks like sports analysis or cooking footage.
• Dynamic video grounding: It analyzes video at up to 2 FPS within a 32K-token context window and can return actionable timecodes for specific moments.
• Multimodal in-context matching: Users can provide a reference image or video and ask the model to find matching instances across new images and videos without fine-tuning or labeled training data.
• Comparison across media: Given two pieces of media, Mk1 can produce a side-by-side comparison, supporting review and inspection workflows.
• Advanced image reasoning: The model supports pointing, counting, OCR, instrument reading, and structured document extraction, including complex layouts, tables, handwriting, and multilingual content.
• Structured spatial outputs: Mk1 can emit point, box, polygon, track, and clip primitives as first-class outputs, making it easier to feed results into downstream robotics or vision systems.

How to Use Perceptron Mk1

A typical workflow starts by submitting an image, a video, or multiple media inputs along with a prompt that specifies the task. Users can ask for object localization, counting, OCR, event detection, timecode extraction, comparison, or structured document conversion.

For robotics and visual pipeline use, the model can be used to label teleoperation footage, identify task boundaries, detect success or failure, and generate annotations that downstream systems can consume directly.

Use Cases

• Video review and event extraction: Analyze long recordings to identify when a specific action occurs, such as grasp attempts, restocking events, or other task milestones.
• Robotics data annotation: Turn teleoperation footage into supervised labels, action-conditioned annotations, quality scores, or subtask boundaries for training downstream models.
• Visual search and asset tracking: Use a reference image or video to locate matching items across new image sets or video streams.
• Industrial inspection and reading tasks: Read gauges, clocks, dashboards, legacy control panels, and messy text in operational environments.
• Document structuring: Convert complex documents into HTML, JSON, or Markdown while preserving layout, tables, hierarchy, and handwritten annotations.

FAQ

Does Mk1 require fine-tuning for matching or detection tasks? No. The source says it can perform in-context matching from a single reference image or video without fine-tuning, a labeled dataset, or a training pipeline.

What kinds of outputs can it produce? It can return text as well as structured spatial outputs such as points, boxes, polygons, tracks, clips, and timecodes, depending on the task.

Is Mk1 only for video? No. The source describes it as strong in image reasoning as well as video and embodied reasoning.

Can it handle long video? It supports dynamic frame-rate analysis at up to 2 FPS within a 32K-token context window, which indicates support for longer-form video analysis, though the source does not state a hard maximum video length.

Alternatives

• General frontier multimodal models: The source compares Mk1 with models from Google, OpenAI, Anthropic, and Alibaba that also handle image and video reasoning, though their output formats and pricing may differ.
• Open-source vision-language models: These may be preferable when teams want open weights or local control, but the source positions Mk1 as a closed-source option focused on performance and structured outputs.
• Robotics perception pipelines with separate components: Some teams may use separate models for detection, OCR, tracking, and annotation, whereas Mk1 aims to combine these steps into one model call.
• Traditional document OCR/extraction tools: These can work well for text-heavy documents, but Mk1 is described as handling more complex layouts, handwriting, and multimodal reasoning in the same workflow.