MJev-Spatial Hits 65.64% on Spatial Benchmarks, 2.5x Faster

MJev-Spatial, a vision-language model built on Qwen3.5-4B, answers spatial questions by scoring candidates with the backbone's pretrained query and key projections instead of generating text, and reports 65.64% overall across twelve spatial benchmarks at 160.6 milliseconds per query on an NVIDIA B200.

The pitch, laid out in a paper posted to Hugging Face, is that spatial judgments at interaction rates (which relation holds, how far an object is, where to place it) have compact output spaces, so autoregressive decoding is overkill. "each classification decision takes a single forward pass and generates no answer tokens," the abstract states. Candidates are either the given options for multiple-choice tasks, or value bins and image regions refined through coarse-to-fine classification for numerical and point prediction.

Trained on 12.3 million question-answer pairs converted from existing spatial datasets, MJev-Spatial "ranks first on seven of twelve benchmarks" and beats its own unmodified Qwen3.5-4B base on all twelve, with gains of 22.89 percentage points on RoboSpatial VQA and 12.50 on RefSpatial-Bench. The abstract reports the model running 2.5× faster than the fastest autoregressive baseline and 5.6× faster than its base, with training throughput up 30.8% from length-aware global batching and cross-GPU load balancing.

It lands as the latest entry in our computer-vision tracker, which has logged 39 stories over the last 90 days.