rayruiyang/VST-7B-SFT

VISIONPricing:Input $0.4 / Output $0.8Concurrent Unit Cost:1Model Size:7BQuant:FP8Context Size:32kPublished:Nov 9, 2025License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

VST-7B-SFT is a 7 billion parameter Vision-Language Model (VLM) developed by rayruiyang, designed to enhance visuospatial abilities. This model, with a 32768 token context length, is part of the Visual Spatial Tuning (VST) framework, which includes VST-P (4.1M spatial perception samples) and VST-R (135K spatial reasoning samples). It achieves state-of-the-art results on spatial benchmarks like MMSI-Bench (32.0%) and VSIBench (60.6%) through a progressive training pipeline combining supervised fine-tuning and reinforcement learning, making it ideal for applications requiring human-like spatial understanding and reasoning.

Loading preview...

Overview

VST-7B-SFT is a 7 billion parameter Vision-Language Model (VLM) developed by rayruiyang, specifically engineered to cultivate human-like visuospatial abilities, encompassing spatial perception and advanced reasoning. This model is part of the comprehensive Visual Spatial Tuning (VST) framework, detailed in the paper "Visual Spatial Tuning".

Key Capabilities

  • Enhanced Spatial Perception (VST-P): Trained on 4.1 million samples across 19 skills, covering single images, multi-image scenarios, and videos, significantly boosting VLMs' spatial perception.
  • Advanced Spatial Reasoning (VST-R): Utilizes 135,000 curated samples to teach models step-by-step and rule-based spatial reasoning.
  • Progressive Training Pipeline: Employs a two-stage approach, starting with supervised fine-tuning for foundational spatial knowledge, followed by reinforcement learning to solidify spatial reasoning.
  • State-of-the-Art Spatial Performance: Achieves strong results on spatial benchmarks, including 32.0% on MMSI-Bench and 60.6% on VSIBench, without compromising general VLM capabilities.
  • Foundation for Physically Grounded AI: Strengthens spatial tuning in Vision-Language-Action Models, paving the way for more robust real-world AI applications.

Good For

  • Applications requiring precise spatial understanding and reasoning from visual inputs.
  • Tasks involving object localization, relative positioning, route planning, and approximate ordering.
  • Developing physically grounded AI systems and robotics where spatial awareness is critical.
  • Researchers and developers looking for VLMs with strong performance on complex spatial benchmarks.

Finetunes
1 models