Perceptron model releases

Perceptron (Perceptron AI, Inc.) is a U.S. AI company based in Bellevue, Washington, founded in November 2024 by Armen Aghajanyan and Akshat Shrivastava, who previously led multimodal research at Meta's FAIR. It builds a family of perception and embodied-reasoning models for the physical world — factory floors, robotics and other physical agents — spanning the open-weight Isaac series and the closed-source Mk (Mark) line. Backed by ~$21M from Bessemer Venture Partners, Foundation Capital, S32 and SmartGateVC. First tracked model is Perceptron Mk1.5 (Sep 2026). This page collects the lab's model releases, lifecycle events, source links, and model metadata in one crawlable record.

Models

Labs

Open

Recent

All models·Latest releases·Open-weight models·Changelog·Home

Most recent in this set

Sep 25, 2026

Perceptron Mk1.5

Perceptron

Sort

1 model

Perceptron Mk1.5

Available

PerceptronProprietary

Perceptron Mk1.5 is Perceptron's embodied-reasoning / perception model for physical agents, released 2026-09-25 as a hosted, proprietary model succeeding Mk1 (May 2026) and retiring the earlier open-weight isaac-0.1/0.2 checkpoints on launch. Multimodal understanding: it accepts text, image, video and audio (WAV/MP3/FLAC) input and answers in natural-language text, optionally emitting machine-parseable spatial/temporal annotations — points, bounding boxes, polygons, tracks and video clips — rather than generating media. Supports graded reasoning via a reasoning_effort control, function/tool calling and JSON-Schema structured outputs. 36,864-token context, 8,192 max output tokens. Pricing $0.15 / 1M input ($0.0375 cached), $1.50 / 1M output. Built by ex-Meta FAIR founders for perception on factory floors and robotics; parameter count undisclosed.

—Undisc.37K ctxSep 25, 2026