README.md

metadata

title: MJev-Spatial
emoji: 🧭
colorFrom: yellow
colorTo: blue
sdk: gradio
sdk_version: 6.30.0
app_file: app.py
short_description: Spatial VLM that decides in one forward pass
python_version: '3.12'
startup_duration_timeout: 30m

MJev-Spatial

Demo of MJev-Spatial: System One Spatial Intelligence via Native Attention Matching (arXiv 2610.22276 · GitHub).

MJev-Spatial is a 4B multimodal decision model built on Qwen3.5-4B. It turns a spatial question into a finite candidate-selection problem and scores those candidates in a single forward pass with no generated answer tokens, reusing the language backbone's final full-attention layer's native query/key projections as the readout.

The demo exposes the three answer spaces from the paper:

• Choice — pick one of the supplied options (e.g. relative-direction questions).
• Scalar — estimate a nonnegative measurement in meters via an interval hierarchy.
• Pointing — point at a requested target in the image; the predicted point is drawn on the image (the model makes three 3×3 region decisions, refining a crop).

Example image: examples/images/scene.png from the official MJev-Spatial repository (Apache-2.0).