SpatialAxiom pushes open spatial reasoning forward

SpatialAxiom makes spatial reasoning more practical for developers by pairing strong benchmark claims with familiar Qwen-based tooling. Its real test will be whether performance transfers from curated spatial evaluations to reliable perception in embodied systems.

• –Two model sizes cover local experimentation and higher-capacity serving
• –Full-parameter SFT keeps the architecture simple and provides a clean base for downstream fine-tuning or reinforcement learning
• –Training spans indoor scenes, egocentric views, and multi-camera settings
• –Reported results lead across several spatial benchmarks, but independent replication remains important
• –The CC BY-NC 4.0 license limits unrestricted commercial deployment