rsoohyun/SpatialBlock-3B-direct
VISIONPricing:Input $0.32 / Cached $0.016 / Output $1.6Concurrent Unit Cost:1Model Size:3BQuant:BF16Context Size:32kPublished:Sep 3, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold
SpatialBlock-3B-direct is a 3 billion parameter vision-language model developed by rsoohyun, fine-tuned from Qwen2.5-VL-3B-Instruct. It specializes in spatial reasoning tasks, directly predicting answers for challenges like 3D-to-2D projection, viewpoint transformation, and structural combination. With a 32768 token context length, this model is optimized for understanding and solving complex spatial problems based on visual input.
Loading preview...
SpatialBlock-3B-direct Overview
SpatialBlock-3B-direct is a 3 billion parameter vision-language model (VL-LLM) developed by rsoohyun, specifically designed to enhance spatial intelligence. It is a fine-tuned version of the Qwen2.5-VL-3B-Instruct model, trained on the synthetic SpatialBlock-15k dataset.
Key Capabilities
- Direct Spatial Reasoning: The model directly predicts answers to complex spatial reasoning tasks without requiring intermediate steps.
- 3D-to-2D Projection: Capable of understanding and processing transformations from three-dimensional objects to two-dimensional representations.
- Viewpoint Transformation: Excels at tasks involving changes in perspective or viewpoint.
- Structural Combination: Proficient in analyzing and predicting outcomes related to combining different structures.
- Vision-Language Integration: Leverages its base as a VL-LLM to interpret visual inputs for spatial problem-solving.
Good For
- Research in Spatial AI: Ideal for researchers exploring methods to improve AI's understanding of spatial relationships.
- Applications requiring direct spatial inference: Suitable for use cases where immediate, accurate answers to spatial queries are needed.
- Benchmarking spatial intelligence: Can serve as a strong baseline for evaluating spatial reasoning capabilities in VL-LLMs.