remyxai/SpaceThinker-Qwen2.5VL-3B
SpaceThinker-Qwen2.5VL-3B is a 3.75 billion parameter multimodal vision-language model (VLM) developed by remyxai, fine-tuned from UCSC-VLAA/VLAA-Thinker-Qwen2.5VL-3B. This model specializes in enhancing quantitative spatial reasoning, including distances, sizes, and grounded object relations, by utilizing test-time compute and synthetic reasoning traces. It excels in tasks requiring precise interpretation of 3D scene context and human-centric spatial logic, making it suitable for embodied AI applications like robotics and drones.
Loading preview...
SpaceThinker-Qwen2.5VL-3B: Enhanced Spatial Reasoning VLM
SpaceThinker-Qwen2.5VL-3B is a 3.75 billion parameter multimodal vision-language model (VLM) developed by remyxai, built upon the Qwen2.5-VL-3B architecture. It is fine-tuned from UCSC-VLAA/VLAA-Thinker-Qwen2.5VL-3B using LoRA (Low-Rank Adaptation) to significantly improve quantitative spatial reasoning capabilities.
Key Capabilities
- Enhanced Quantitative Spatial Reasoning: Excels at inferring accurate distances, sizes, and object properties within 3D scenes.
- Grounded Object Relations: Interprets spatial relationships like "left-of," "above," and "inside" with consistent explanations.
- Test-Time Compute Integration: Utilizes reasoning traces generated by the VQASynth pipeline to ground responses on scene observations.
- Robust Performance: Demonstrates strong results on benchmarks like OmniSpatial, SpatialScore (especially the Hard split), and SpaCE-10, often outperforming larger models or achieving comparable results to frontier APIs when prompted for step-by-step reasoning.
Use Cases
- Embodied AI: Critical for applications demanding spatial understanding, such as robotics and drones, for planning and navigation.
- Visual Question Answering (VQA): Particularly effective for VQA tasks requiring precise spatial and quantitative answers.
- Research & Development: Provides a strong baseline for further research into spatial reasoning in VLMs, especially with its detailed dataset and training methodology.
SpaceThinker-Qwen2.5VL-3B was trained on the ~12K sample SpaceThinker dataset, which includes synthetic spatial reasoning traces covering various question types like spatial relations and distances. While powerful, users should be aware of potential performance degradation in cluttered environments and inherent multimodal biases from the base model.