SpatialAxiom/SpatialAxiom-35B-A3B

SpatialAxiom-35B-A3B: An Open Spatial Intelligence Model

SpatialAxiom-35B-A3B is a 35.1 billion parameter Mixture-of-Experts (MoE) vision-language model, featuring 3 billion active parameters. Developed by SpatialAxiom, this model is built upon the Qwen3.5 VLM architecture, maintaining a general-purpose multimodal design without task-specific architectural modifications. Its strength is derived from a unique spatial data-centric training recipe, which includes a systematic taxonomy of spatial tasks, balanced task distribution, and high-quality data synthesis.

Key Capabilities

• Leading Spatial Reasoning: Achieves top performance on benchmarks such as VSI-Bench, MMSI-Bench, MindCube, ViewSpatial, and EmbSpatial, often surpassing larger proprietary and open-source models.
• 3D Relational Inference: Excels at understanding and inferring relationships in 3D spaces.
• Perspective Taking: Capable of interpreting scenes from different viewpoints.
• Multi-view Correspondence: Handles information from multiple visual perspectives.
• Embodied Video Understanding: Processes and understands spatial dynamics within video content.
• High Context Length: Supports a native context length of 262,144 tokens, extensible up to 1,010,000 tokens.
• Full-Parameter SFT: Trained purely with full-parameter Supervised Fine-Tuning, providing a clean base for further post-training or downstream fine-tuning.

Good For

• Applications requiring advanced spatial intelligence and reasoning.
• Research and development in 3D scene understanding, robotics, and embodied AI.
• Developers looking for a robust VLM backbone with strong spatial capabilities for further fine-tuning or reinforcement learning.
• Use cases involving image and video analysis where spatial relationships are critical, such as analyzing room layouts or understanding routes in videos. Note that it is fine-tuned for direct responses only and does not support thinking blocks or chain-of-thought modes.