Spatial understanding is a fundamental cornerstone of human-level intelligence. Nonetheless, current research predominantly focuses on domain-specific data production, leaving a critical void: the absence of a principled, open-source engine capable of fully unleashing the potential of high-quality spatial data. To bridge this gap, we elucidate the design principles of a robust data generation system and introduce OpenSpatial -- an open-source data engine engineered for high quality, extensive scalability, broad task diversity, and optimized efficiency. OpenSpatial adopts 3D bounding boxes as the fundamental primitive to construct a comprehensive data hierarchy across five foundational tasks: Spatial Measurement (SM), Spatial Relationship (SR), Camera Perception (CP), Multi-view Consistency (MC), and Scene-Aware Reasoning (SAR). Leveraging this scalable infrastructure, we curate OpenSpatial-3M, a large-scale dataset comprising 3 million high-fidelity samples. Extensive evaluations demonstrate that versatile models trained on our dataset achieve state-of-the-art performance across a wide spectrum of spatial reasoning benchmarks. Notably, the best-performing model exhibits a substantial average improvement of 19 percent, relatively. Furthermore, we provide a systematic analysis of how data attributes influence spatial perception. By open-sourcing both the engine and the 3M-scale dataset, we provide a robust foundation to accelerate future research in spatial intelligence.
Figures & tables
Figure 1: Overview of OpenSpatial. The left panel provides a high-level schematic of the OpenSpatial pipeline. The right panel demonstrates that models trained with OpenSpatial-generated data exhibit significant improvements in spatial intelligence. Notably, the evaluation benchmarks are consistent with those reported in Tab. 1 .
Figure 2: Illustration of the data engine. The left panel of the figure illustrates the data processing and annotation pipeline, while the right panel presents the detailed statistics of the dataset, including source data distribution and task distribution.
Figure 3: Overview of the OpenSpatial dataset. OpenSpatial-3M comprises 3M high-quality samples for spatial understanding, encompassing five primary categories: Spatial Measurement (SM), Spatial Relationship (SR), Camera Perception (CP), Multi-view Consistency (MC), and Scene-Aware Reasoning (SAR). Representative cases for each category are curated and illustrated above. For brevity, the displayed QA pairs have been condensed; please refer to the Appendix for comprehensive details.
Models
3D-Avg
BLINK
AllAngles
ERQA
VSI
3DSR
MMSI
CV-3D
RealWorldQA
MMStar
MMB
MMU
Proprietary Models
Gemini-2.5-Pro [ 14 ]
62.4
70.6
61.3
55.8
48.4
57.6
36.9
91.3
77.3
77.5
90.1
81.7
Open-source General Models
InternVL2.5-4B [ 61 ]
47.3
50.8
45.1
41.0
28.3
44.0
28.5
76.4
64.2
58.5
78.5
50.0
InternVL2.5-8B [ 61 ]
51.6
54.9
48.9
40.8
39.3
51.0
28.6
79.9
69.4
62.6
82.3
53.3
InternVL3-2B [ 61 ]
47.9
52.8
48.6
36.3
30.3
46.4
25.9
77.3
65.5
61.5
77.7
45.9
Table 1: Performance comparison of representative VLMs on spatial reasoning and general multimodal benchmarks. Dark blue and light blue backgrounds denote the best and second-best results, respectively. All models are evaluated under the same codebase to ensure a fair comparison. Note : 1)VSI is tested with 16 frames. 2)‘ ± ’ denotes the performance gain or loss relative to the baseline model.
Data source
Data size
MAD
Std. Dev.
BLINK
AllAngles
ERQA
VSI
3DSR
MMSI
CV-3D
RealWorldQA
Cambrain-S [ 54 ]
590k
-6.0
5.4
54.1( -10.1 )
48.6 ( -5.3 )
39.2( -3.0 )
57.0
49.7( -2.3 )
29.2( -7.1 )
75.3( -17.9 )
68.1( -2.6 )
SenseNova-SI [ 7 ]
800k
-6.5
7.0
59.7 ( -4.5 )
47.5( -6.4 )
38.0( -4.2 )
55.0( -2.0 )
48.9 ( -5.1 )
36.3
69.0( -24.2 )
65.1( -5.6 )
VST [ 53 ]
500k
-2.8
3.9
61.4( -2.8 )
50.7( -3.2 )
42.2
44.6( -12.4 )
54.1
32.4( -3.9 )
93.2
70.6( -0.1 )
OpenSpatial(subset)
500k
-2.5
4.4
64.2
53.9
41.3( -0.9 )
43.0( -14.0 )
52.0( -2.1 )
34.7( -1.6 )
91.8( -1.4 )
70.7
Table 2: Comparison with open-source datasets for spatial reasoning. ‘ − ’ indicates the difference from the best result . MAD: Mean Deviation; Std. Dev.: Standard Deviation. Note : The subset of OpenSpatial is constructed independently and does not include any data from SenseNova-SI.
Setting: data size(200k)
BLINK
AllAngles
VSI
CV-3D
Qwen2.5-VL-7B
55.3
50.1
36.0
73.8
+Point Cloud Centric
57.2
49.7
37.2
83.7
+3D-Box Centric
60.3
53.2
41.7
89.9
+3D-Box Centric (no filter)
56.6
47.0
32.1
78.2
Table 3: Rationality Analysis of Module Design.
Data Size
3D-Avg
BLINK
AllAngles
ERQA
VSI
3DSR
MMSI
CV-3D
RealWorldQA
20%
57.6
64.9
55.0
41.0
51.8
52.0
34.7
91.8
69.5
40%
58.5
66.2
56.8
42.1
52.4
52.9
35.8
91.0
71.1
60%
58.6
66.9
56.7
42.0
53.2
52.3
36.0
92.1
70.7
80%
59.2
66.2
59.7
41.2
53.5
53.0
37.6
91.5
70.5
Full
59.7
65.9
58.4
41.8
56.7
54.3
39.6
92.5
68.3
Table 4: Ablation on data scaling. Color Map : worst to best .
Model Size
3D-Avg
BLINK
AllAngles
ERQA
VSI
3DSR
MMSI
CV-3D
RealWorldQA
OpenSpatial-Qwen2.5-VL-3B
56.1
61.0
53.0
40.0
55.2
47.8
32.0
94.0
66.0
OpenSpatial-Qwen2.5-VL-7B
59.7
65.9
58.4
41.8
56.7
54.3
39.6
92.5
68.3
OpenSpatial-Qwen2.5-VL-32B
61.3
68.2
63.3
44.0
57.3
55.3
39.8
93.4
69.3
Table 5: Ablation on model scaling. Dark blue : best result. Light blue : second best.
Figure 4: Visualization of 3D lifting results from in-the-wild outdoor web data.
Figure 5: Impact of diverse tasks on spatial intelligence. Best zoomed in.
3D spatial reasoning studies how AI systems infer geometry, object relations, viewpoints, and motion to understand and act in three-dimensional scenes.