BELLEVUE, Wash.: Perceptron AI today announced the launch of its model purpose-built for video understanding and embodied reasoning. It delivers performance competitive with leading frontier models – including Google, Anthropic, OpenAI, and Qwen – at a fraction of the cost.

Perceptron Mk1 (“Mark One”) achieves competitive scores across leading image, video, and spatial-reasoning benchmarks, matching or exceeding other frontier models while operating at the cost of lightweight alternatives. Organizations that previously had to choose between capability and cost can now deploy high-accuracy video understanding at scale.

“We built Perceptron to make the physical world legible to AI systems,” said Armen Aghajanyan, Co-founder & CEO. “Until now, frontier visual understanding has come with a cost that’s out of reach for most industrial and consumer applications. We’ve changed that.”

Built for the Real World

The model is designed to close the loop between digital intelligence and physical action. It is built to work everywhere reliable visual understanding is business-critical:

• Manufacturing & Industrial: Operational and safety analytics across factories, construction sites, and warehouses. Detect product defects, OSHA violations, read analog instruments on inspection rounds, and track inventory levels.
• Media & Content: Semantic visual search, tagging, and policy enforcement. Clip sports highlights, search film and TV libraries, moderate AI-generated content.
• Robotics & Automation: Onboard embodied reasoning for manipulation and navigation: pointing for grasps, multi-view understanding across cameras, and success detection. Plus, offline curation of teleoperation episodes into training data.
• Geospatial & Critical Infrastructure: Satellite, drone, and fixed-camera analysis. Detect vegetation encroachment, oil rig and bridge anomalies, construction progress, and disaster damage for claims.
• Security & Surveillance: Real-time intelligence across home and enterprise cameras. Surface context-aware alerts that separate meaningful events from background activity.
• Devices & Agentic Tooling: Enhanced vision for Claude, Codex, and other text-first agents. Assess documents, sort files, and automate desktop tasks, like browser use.

"Robotics is the hardest test of real-world, physical AI," said Akshat Shrivastava, Co-founder & CTO. "It demands perception, reasoning, and action in one closed loop, under real-world conditions. Solve that and every other physical AI problem becomes tractable. That's the bar we built Mk1 to clear.”

Availability

Perceptron’s model is available today through the Perceptron AI API platform and via OpenRouter, enabling developers and enterprises to integrate visual intelligence and embodied reasoning directly into existing systems. To get started, try the model on the Perceptron Demo Application.

About Perceptron AI

Perceptron AI is an AI research lab pushing the frontier of physical AI, building systems that perceive, reason, and act in the real world, reliably and at scale. Headquartered in Bellevue, Washington, Perceptron was founded in November 2024 by Armen Aghajanyan and Akshat Shrivastava, both former research scientists at Facebook AI Research (FAIR).

Fonte: Business Wire