Perceptron AI releases Perceptron Mk1 multi-modal model
Perceptron AI released Perceptron Mk1, a multi-modal model for frontier video understanding and embodied reasoning. The model was developed over 16 months with custom recipes optimized for physical-world performance. Demonstrations show it extracting highlights from overhead soccer footage and generating robotic arm optimization recommendations amid factory obstructions. The release is the company's first public model and is distributed via Hugging Face.
@ArmenAgha It was one of the main/coolest features of OWL-ViT quite a while ago. Luckily you said almost all 😁
Armen Aghajanyan@ArmenAghaPointing by example is a surprisingly useful capability almost all multimodal models except ours get wrong.
#437Armen Aghajanyan@ARMENAGHAQuick heads up: Mk1 is seeing much higher API traffic than we forecasted, so some users may see intermittent throughput regressions over the next few days. We’re adding capacity aggressively and will keep pushing on reliability. Thanks for bearing with us.
Armen Aghajanyan@ArmenAghaI'm excited to finally release the fruit of the research we've been doing at Perceptron for the last 16 months: Perceptron Mk1. We've been developing multi-modal recipes from the ground up to build models that perform best in the physical world, from video understanding to embodied reasoning to robotics. Mk1 is our scaled up recipe.
#437Armen Aghajanyan@ARMENAGHAPointing by example is a surprisingly useful capability almost all multimodal models except ours get wrong.
#437Armen Aghajanyan@ARMENAGHAI'm excited to finally release the fruit of the research we've been doing at Perceptron for the last 16 months: Perceptron Mk1. We've been developing multi-modal recipes from the ground up to build models that perform best in the physical world, from video understanding to embodied reasoning to robotics. Mk1 is our scaled up recipe.
Perceptron AI@perceptronincToday we're releasing Perceptron Mk1: frontier video and embodied reasoning.
Mk1 is incredibly cost effective. It performs at par with Gemini-Flash/Gemini-ER as well as the larger open source Qwen models on all perceptive/physical AI tasks, at a fraction of the cost ($0.15/M input, $1.50/M output).
Armen Aghajanyan@ArmenAghaI'm excited to finally release the fruit of the research we've been doing at Perceptron for the last 16 months: Perceptron Mk1. We've been developing multi-modal recipes from the ground up to build models that perform best in the physical world, from video understanding to embodied reasoning to robotics. Mk1 is our scaled up recipe.
On Embodied Reasoning tasks, Mk1 hits the frontier while running cheaper and faster.
Armen Aghajanyan@ArmenAghaMk1 is incredibly cost effective. It performs at par with Gemini-Flash/Gemini-ER as well as the larger open source Qwen models on all perceptive/physical AI tasks, at a fraction of the cost ($0.15/M input, $1.50/M output).
Armen Aghajanyan@ArmenAghaOn Embodied Reasoning tasks, Mk1 hits the frontier while running cheaper and faster.
Video understanding is best in class.
Armen Aghajanyan@ArmenAgha
Armen Aghajanyan@ArmenAghaVideo understanding is best in class.
Try it: http://demo.perceptron.inc
Interested in the weights? We'll be opening a small partners program for folks to get direct access to the model. DM me.
Armen Aghajanyan@ArmenAgha#437Armen Aghajanyan@ARMENAGHAWe're throwing a release party in Bellevue, lined up with MLSys. Thursday May 21, 6 to 9pm; food, drinks, RSVP at https://partiful.com/e/bsC2wZxXPhFiDbMynb5l
Armen Aghajanyan@ArmenAghaI'm excited to finally release the fruit of the research we've been doing at Perceptron for the last 16 months: Perceptron Mk1. We've been developing multi-modal recipes from the ground up to build models that perform best in the physical world, from video understanding to embodied reasoning to robotics. Mk1 is our scaled up recipe.
@ArmenAgha @Scobleizer Congrats, very cool model!
Armen Aghajanyan@ArmenAghaI'm excited to finally release the fruit of the research we've been doing at Perceptron for the last 16 months: Perceptron Mk1. We've been developing multi-modal recipes from the ground up to build models that perform best in the physical world, from video understanding to embodied reasoning to robotics. Mk1 is our scaled up recipe.
#1754alex peysakhovich@ALEX_PEYSnew architecture ideas coming out of ai startups are cool as hell
Armen Aghajanyan@ArmenAghaI'm excited to finally release the fruit of the research we've been doing at Perceptron for the last 16 months: Perceptron Mk1. We've been developing multi-modal recipes from the ground up to build models that perform best in the physical world, from video understanding to embodied reasoning to robotics. Mk1 is our scaled up recipe.