VLX-Seek 1.5: embodied perception in 0.6B/3B/10B, fewer illusions

Om AI Lianhui released VLX-Seek 1.5 on July 6 — a fine-grained perception VLM new version for on-device embodied scenarios. The keyword for this update is not "bigger" but "more deployable". The new version plans 0.6B / 3B / 10B three-tier model series, so that drones, robot dogs, surveillance cameras, and other terminals with different compute budgets can pick a suitable version — a rare engineering awareness in the era of mindlessly scaling VLMs that ultimately can't fit on-device. The architecture introduces more Linear Attention layers and faster OPN candidate region generation, with inference latency more friendly to the device. More noteworthy is the handling of "object hallucination". In embodied scenarios, the cost of a robot incorrectly tracking a non-existent target is far greater than missing one. VLX-Seek 1.5 introduces an explicit None output format: when the user asks about A and B in the image, if B doesn't exist, the model must output A's coordinates + B's None. On three benchmarks (HumanRef, VisDrone, RefDrone), the Object Hallucination metric (FP / GT count) is lower than both the previous version and LocateAnything. On visual capability, the new training data adds more drone, surveillance, and robot viewpoints, and the auxiliary vision tower is also upgraded. On COCO, LVIS, RefCOCO, VisDrone, RefDrone, EmbSpatialBench and other benchmarks, VLX-Seek 1.5-3B surpasses several larger open-source/closed-source VLMs. Om AI Lianhui announced the open-sourcing of the 10B version. In the embodied-perception track, this is a rare "10B can also run locally" open-weight foundation — for robot/drone developers, no longer having to compromise performance just to "fit on device" is finally a reality.