LVL-1: A 100M Parameter Micro-World Model

A useful description of a physical scene must preserve several kinds of evidence at once: what is present, where it appears, how surfaces and objects are arranged, and what interaction is unfolding.

Object identity is semantic knowledge. Boundaries, parts, and relative position are local and dense knowledge. Relations such as above, inside, held by, or approachingrequire those visual signals to survive long enough to become language.

LVL-1 is an experiment in extracting those specific priors from specific models. It is not a metric 3D reconstruction and it does not maintain a world state through time. It turns one image into a compact, local observation that another spatial system can retain, compare, or reject.

The <DESCRIBE/> task rewards visible objects, spatial relations, activity, and scene context. The<2D_BBOX/> task asks the same model to retain explicit image-space localization. They share the visual prefix but demand different outputs.

No-image and wrong-image counterfactuals are equally important. They penalize a decoder that can imitate familiar training prose without consulting the current image. Text replay protects the language core from collapsing into a narrow caption template.

The encoder names do not guarantee grounding. Spatial knowledge only reaches the output when the projector, token budget, supervision, and counterfactual tests all preserve it.

LVL-1 can produce useful descriptions of robotics workcells, electronics benches, tools, containers, people, and everyday environments. It also has clear limitations.

The model can repeat objects, miss fine details, hallucinate plausible elements, or lose grounding outside its training distribution. It is not a safety system, a geometric measurement tool, or a robot controller.

The next work is better spatial resolution, cleaner grounded supervision, less repetition, and evaluation that measures whether the model is truly using the image.

Vitrus is building machines that need to observe and operate in changing physical environments. We do not expect one small model to understand an entire factory. We expect many specialized observers to contribute evidence to a shared spatial system.

LVL-1 describes what a camera can see. Calibration and geometry establish where things are. VitrusOS maintains state, validates constraints, and controls what machines are allowed to do.

A concrete point of reference is Cosmos-Reason1-7B, which consumes video and text and is trained for physical common sense and embodied reasoning. NVIDIA'sCosmos-Predict2.5 instead uses 2B and 14B models to predict future world states as video. Those are fundamentally broader tasks than LVL-1's single-image extraction.

The systems can be complementary. A Cosmos model can serve as a teacher, reasoner, or simulator upstream; a compact model like LVL-1 can run beside cameras and machines, continuously turning familiar visual scenes into low-cost evidence. The comparison above is architectural—not a quality or speed leaderboard. There is no shared checkpoint, task, hardware, or benchmark that would support a direct score.