---

license: apache-2.0

base_model:

• rsoohyun/SpatialBlock-3B-direct

language:

• en

pipeline_tag: image-text-to-text

library_name: transformers

tags:

• text-generation-inference
• llama-cpp

---

SpatialBlock-3B-direct-GGUF

> SpatialBlock-3B-direct is an open-source multimodal model released on Hugging Face by rsoohyun under the Apache-2.0 license, developed to enhance spatial intelligence in Large Vision-Language Models (LVLMs). Based on the Qwen/Qwen2.5-VL-3B-Instruct base architecture and supported by the Hugging Face transformers library via the image-text-to-text pipeline, this checkpoint is fine-tuned on the synthetic SpatialBlock-15k dataset as presented in the paper SpatialBlock: Enhancing Spatial Intelligence in LVLMs via Synthetic Block-Stacking Problem. It specializes in directly predicting solutions to intricate spatial tasks—including 3D-to-2D projection, viewpoint transformation, and structural combination—with complete training methodology, evaluation metrics, and the companion “reason” model accessible via its GitHub repository.

Model Files

File Name | Quant Type | File Size | File Link |

|-----------|------------|-----------|-----------|

| SpatialBlock-3B-direct.BF16.gguf | BF16 | 6.18 GB | Download |

| SpatialBlock-3B-direct.Q4_K_M.gguf | Q4_K_M | 1.93 GB | Download |

| SpatialBlock-3B-direct.Q5_K_M.gguf | Q5_K_M | 2.22 GB | Download |

| SpatialBlock-3B-direct.mmproj-bf16.gguf | mmproj-bf16 | 1.34 GB | Download |

llama.cpp

LLM inference in C/C++ — https://github.com/ggml-org/llama.cpp