Perceptron API: Vision-Language Models for Physical AI

The Perceptron API serves all of Perceptron's vision-language models through a single OpenAI-compatible interface. One integration covers video understanding, embodied reasoning for robotics, agentic object detection, OCR, multimodal search, and low-latency perception at the edge.

Credits will be allocated to your Perceptron account and can be used across all models available on the Perceptron Platform.

For any questions, please contact: support@perceptron.inc 

PERCEPTRON MODELS

Perceptron Mk1: Perceptron's highest-quality vision-language model for video and embodied reasoning. It matches or exceeds the performance of flagship models from other frontier labs at an order of magnitude lower cost per token.

• Accepts image and video inputs paired with natural language queries and produces detailed visual understanding responses in either structured or natural language.
• Excels at video QA, summarization, event detection, and video clipping with start/end timestamps.
• On images: point-by-example grounding from multimodal prompts, OCR and document parsing on messy real-world inputs, open-vocabulary object detection and counting, and hand pose estimation.
• Reasoning can be enabled per request to trade latency for deeper analysis on harder tasks. Structured annotations are emitted inline when explicitly requested.
• In-context learning: teach new visual concepts with a handful of annotated image or video examples, without fine-tuning.

ISAAC MODELS (OPEN WEIGHTS)

Isaac 0.2 (2B): Best-in-class open-weight 2B vision-language model with reasoning. Sub-200 ms time-to-first-token for fast image understanding.

Isaac 0.2 (1B): Compact 1B vision-language model with reasoning, optimized for edge and low-latency deployments.

Isaac 0.1: Original 2B vision-language model, fully supported for existing integrations.

AGENTIC APIs

Perceptron Agentic Detection: Frontier performance on dense, ambiguous detection tasks.

Describe your detection classes in plain language or with visual examples. An agent coordinates multiple Mk1 calls, zooming in to resolve the small, dense, and ambiguous details that one-shot detectors miss at full-image resolution.

One detection API that generalizes across satellite, drone, robot, vehicle, and wearable imagery.

MULTIMODAL PRODUCTS

Multimodal Search: Search image and video libraries with natural language or a reference example.

Every model is served through the same OpenAI-compatible API, with structured outputs (JSON Schema, Pydantic, or regex) and a Python SDK with grounded perception primitives (points, boxes, polygons, and clips).

Perceptron models are used in physical AI applications, including robotics, manufacturing, logistics, retail, security, media, and geospatial.

Resources: https://docs.perceptron.inc