Perceptron Mk1.5: an embodied model that answers in coordinates

Structured geometry a robot controller can read without a second parser, on OpenRouter now — plus Perceptron kills its Isaac IDs and Meta hands out Ubuntu VMs.

Nowline SEP 27 8:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • It replies in points, boxes, polygons and tracks

    Mk1.5 takes text, image, video and audio and returns machine-parseable geometry next to its text — bounding boxes, polygons, keypoints, and `<track>` sequences that follow an object across a whole video. A robot controller or QA pipeline can read the coordinates directly, with no second parsing stage.

  • Cheap, fast, and callable on OpenRouter today

    It's live on OpenRouter and Perceptron's own API at $0.15 / $1.50 per million tokens in/out ($0.0375 cached), with a 37K context and 8K max output. Perceptron claims up to 4.7x faster short answers than Mk1 and 0.94 vs Gemini 3.8 Flash's 0.62 on egocentric hand-box tracking — all vendor numbers, no third-party arena scores yet.

  • Isaac's model IDs are gone — check your error handling

    Same day, Perceptron retired isaac-0.1, isaac-0.2-1b and isaac-0.2-2b-preview; those IDs now return 404. Unknown model IDs also switched from 400 to 404, and the old `vision_config` boolean is replaced by a `reasoning_effort` knob (high/medium/low/minimal/none) that defaults to high — watch the per-call cost before you ship.

  • Elsewhere: Meta Muse hands every user an Ubuntu VM

    Meta's Muse agent now gives each user a full cloud computer running Ubuntu on AMD EPYC Turin (2 cores, 8GB), with the agent passing terminal commands straight to the host. It's effectively a free sandbox you can point an agent at — Meta concedes the design was heavily inspired by OpenClaw.