Om AI Lab Ships 10B Vision Model Without Bounding Boxes

Om AI Lab shipped a 10-billion-parameter vision-language model for drones, robots, and surveillance, licensed Apache 2.0, according to its Hugging Face model card.

The pitch is for builders doing visual grounding. Instead of training a model to output pixel coordinates for a bounding box, VLX-Seek-1.5-10B converts image regions into tokens the language model can address directly. The bet: language models are better at pointing to a token than predicting a coordinate, so let the model do what it's actually good at.

The model handles open-vocabulary detection, referring expression comprehension, multi-object grounding, and counting backed by region-level evidence, per the listing. That's the feature set robotics and surveillance teams need in one model: find the thing, describe where it is, count how many, without bolting on a separate detector.

It sat at 349 downloads and zero Reddit comments despite 65 upvotes, drowned out by Meta.

Here's the bet worth watching: if language-addressable region tokens actually outperform raw coordinate regression for grounding tasks, every VLM still training on box-coordinate output is building on the wrong foundation. Nobody's benchmarked that head to head yet. Watch whether downloads and fine-tunes pick up once someone does.

Om AI Lab's model became visible on August 10, the same window Meta pulled attention toward itself and away from smaller open-weight labs.