SpatialAxiom/SpatialAxiom-9B

SpatialAxiom-9B: An Open Spatial Intelligence Model

SpatialAxiom-9B is a 9 billion parameter vision-language model (VLM) developed by D2I-ai, designed for advanced general spatial reasoning. Built upon the Qwen3.5 VLM architecture, it specializes in tasks such as 3D relational inference, perspective taking, multi-view correspondence, and embodied video understanding.

Key Capabilities and Features

• Leading Spatial Reasoning: Achieves top performance on benchmarks like VSI-Bench, MMSI-Bench, MindCube, ViewSpatial, and EmbSpatial, often surpassing larger proprietary and open-source models.
• Spatial Data-Centric Training: Utilizes a systematic taxonomy of spatial tasks, balanced task distribution, and data synthesis to enhance data quality, trained purely with full-parameter Supervised Fine-Tuning (SFT).
• Qwen3.5 VLM Backbone: Inherits the robust Qwen3.5 architecture, maintaining a general-purpose multimodal design without task-specific architectural modifications.
• Multimodal Input Support: Processes both image and video inputs for spatial analysis, with a native context length of 262,144 tokens, extensible up to 1,010,000 tokens.
• Open-Weight Release: Fully compatible with transformers and vLLM for easy integration and deployment.

When to Use SpatialAxiom-9B

• Applications requiring precise spatial understanding: Ideal for robotics, augmented reality, virtual reality, and autonomous navigation.
• Research in spatial AI: Serves as a clean, strong baseline for further fine-tuning or reinforcement learning in spatial intelligence tasks.
• Video analysis for spatial relationships: Excels at interpreting spatial layouts and movements within video content.

Note: SpatialAxiom-9B is fine-tuned for direct responses only and does not support 'thinking blocks' or enable_thinking mode during inference.