I’m back, with a little something… GPC 1, a new general-purpose classifier enabling: bounding boxes, poses, coordinates, direct mechanical control and so much more, all in milliseconds. Post-trained on millions of datapoints and thousands of unique problem classes learned directly from you guys posting on X, GPC-1 is a... great tool for many things. Early user feedback is also good :) A new output class enables precise continuous range numerical outputs, so the model can output coordinates of bounding boxes, exact angles and distances, pose estimations, scores, and more. It goes a lot beyond typical classification which just pick a label. Excited to see where this goes. Available now!show more

i wish i can give everyone a hug but here’s the next best thing 🤗

GPC-1 uses a new dual training technique, jointly training the Qwen base model and add on adapters. As it is, its great, but if you want to customize more, the model was literally trained to be really easy to fine tune for anything 🚀

this is awesome! thx for sharing. i did something similar but went the opposite direction of using a small vlm to run locally, though no boxes or anything.

Love it!

mmmm yes, this won't be used for robotics and anything that contributes to society but used to trade sports prediction books with cameras within stadiums

humanity at its best 😂

you cooked with this one. since the Qwen backbone supports video, curious if you tried GPC-1 on videos?

not yet but good idea!

@yoheinakajima Very nice. Congrats. Now I’m curious as to if I can integrate this and how well it will work for the avatar portion of my twin.

@yoheinakajima Tons of medical data is being training on now

@yoheinakajima What’s the medical data that you’re training on and what’s the goal with that training?

@yoheinakajima goal rn is to optimize hospital scheduling for surgery based on patients (but training data is broad, including data on how to control lab equipment)

if it really operates in milliseconds, that could seriously change automation workflows

It rly does and it rly will, already deploying it in a few labs rn

Super cool, classifiers are having the moment they deserve! Did you try this out on maps or geospatial stuff at all?

Good idea! Will include it in v1.1

Happy to give specific use cases / feedback if & when you get into it

DM me anytime, I’m always into it

Cool, wasm port?

Awesome video Harsha

This is so cool. There's not enough time in the day to try out all the new models people are creating. How much vram does this use?

Congrats Harsha! We’re currently in stealth mode building a humanoid robot 🤖 and would love to learn more about GPC-1. Do you see it being useful for robot control, like watching a human do a task and have a robot replicate the pose and how would you advise using it for this?

woah this was exactly what i hoped jev could do

@yoheinakajima Will this work for multiple objects in video?

@yoheinakajima Yup, just describe the object in the key (i.e person_to_the_left_x_axis, not the most elegant solution rn but working on it)

@yoheinakajima I could see so may things being useful in sport tracking. I’m working on analysing a football training drill from coach’s perspective. I would love to try it

@yoheinakajima lmk how it performs for you!

This is so cool! Whats the latency like?

Depends on what hardware you use, under 100ms on a h100

I'm not sure for robotics because those arms pass through each other, but looks okay for less critical setups

ironing those issues out now

Damn...just an unrelated question...how do you make these prod demo videos?

one shotted with Gemini and Astra

How does this do on OCR tasks on docs?

Trained on many OCR examples, lmk how it performs at your use case

@CDerinbogaz

Bro is too smart!!

Impressive that you got continuous numeric output in ms, could be a game changer for real time robotics pipelines

Yolo with natural language