hunarbatra/SpatialThinker-7B

VISIONPricing:Input $0.4 / Output $0.8Concurrent Unit Cost:1Model Size:7BQuant:FP8Context Size:32kPublished:Sep 15, 2025License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

SpatialThinker-7B by Hunar Batra et al. is a 7 billion parameter 3D-aware multimodal large language model (MLLM) built upon Qwen2.5-VL-7B-Instruct, designed for advanced spatial reasoning. It integrates structured spatial grounding with multi-step reasoning through reinforcement learning, simulating human-like spatial perception by constructing scene graphs and utilizing dense spatial rewards. This model excels at complex spatial understanding tasks, outperforming larger models like GPT-4o on specific benchmarks despite being trained on a smaller dataset of 7,587 spatial VQA samples.

Loading preview...

SpatialThinker-7B: Advanced 3D-Aware Multimodal Reasoning

SpatialThinker-7B is a 7 billion parameter multimodal large language model (MLLM) developed by Hunar Batra and colleagues, specializing in 3D-aware spatial reasoning. Built on Qwen2.5-VL-7B-Instruct, this model employs a novel approach using reinforcement learning with dense spatial rewards to achieve human-like spatial perception. It constructs detailed scene graphs of objects and their spatial relations, enabling multi-step reasoning for complex visual questions.

Key Capabilities

  • Structured Spatial Reasoning: Generates question-focused scene subgraphs, including objects, bounding boxes, and their inter-relations.
  • Dense Spatial Rewards: Utilizes a multi-objective reward function during training (GRPO) to enforce accuracy in format, count, and spatial grounding.
  • Comprehensive Spatial Understanding: Addresses 9 distinct spatial reasoning categories: relations, reach, size, orientation, instance location, depth, distance, count, and existence.
  • High Performance: Demonstrates superior performance over models like GPT-4o on spatial understanding benchmarks, despite being trained on a relatively small dataset of 7,587 spatial VQA samples (STVQA-7K).

Good for

  • Applications requiring precise spatial understanding from images.
  • Research into human-like spatial perception and reasoning in AI.
  • Developing systems that need to interpret complex visual scenes with detailed object relationships.
  • Benchmarking advanced multimodal reasoning capabilities.

For more details, refer to the project page and the arXiv paper.

Finetunes
1 models