After 3B Model Outperforms NVIDIA and Google, Om AI Launches End-Side Native VLX Model: Small Parameters Achieve Precise Perception of Physical World

On August 6, Om AI Lianhui announced a multi-hundred-million yuan financing led by Qianhai Mother Fund, and simultaneously open-sourced VLX-Seek 1.5, described as the world's first end-side native model. The company positions VLX as an 'end-side native' approach, contrasting with cloud-dependent strategies from NVIDIA, Google, and others. VLX-Seek 1.5 comes in 3B and 10B parameter versions and features a streaming multimodal architecture that processes video streams in real-time on-device, unlike traditional VLM batch processing. In benchmarks, the 3B version outperformed NVIDIA's LocateAnything-3B in LVIS Mean (57.5 vs 50.7), RefCOCOg test Mean (80.2 vs 76.8), and RefDrone F1 (73.2 vs 52.3), with significant improvements in small-object and drone-view scenarios. The model also introduces a target hallucination metric, achieving lower false positive rates (FP/GT 18 vs 71.3). The company also highlighted its 'one brain, multiple forms' agent platform OmAgent, OttoPlex for embodied scenarios, OttoBox AI Studio developed with Lenovo and Apple, and Homer AI wearable visual assistant serving nearly 100,000 visually impaired users. The open-sourcing aims to define a standard for end-side native AI in physical world applications.