Why this matters
Robotic perception and interactive vision tasks increasingly demand models that combine broad visual understanding with explicit spatial and affordance reasoning. This release targets that gap by building a 10B-scale multimodal decoder with training and evaluation focused on fine-grained spatial relations, multi-view association, and multi-step agentic workflows.
Key Capabilities
- Spatial and embodied reasoning: trained on tasks covering left/right, occlusion, depth ordering, 2D↔3D rotation, paper-folding and part-motion reasoning, enabling explicit geometric answers and grounding useful for VLA and robotics planning.
- Multimodal inputs at scale: accepts text, single or multiple images, and video; any-resolution visual input and a long 128K-token context window allow cross-view association and event retrieval from long footage.
- Agentic/tool-use readiness: designed to produce multi-step, multi-turn tool-call outputs and structured reasoning; integrates an entropy-gated recurrent refinement mechanism to allocate extra compute to harder tokens.
- Practical integration: packaged examples for vLLM/Docker serving and transformer-based offline inference, with recommended sampling settings for spatial grounding versus general tasks.
Who it's for and trade-offs
Great fit if you need a locally hostable, open-weight multimodal model that prioritizes geometric reasoning, multi-view consistency and agent-style tool planning — e.g., robotics researchers, embodied-AI teams, and vision+agent prototyping groups. Look elsewhere if you need the absolute top performance on non-spatial general VQA/OCR benchmarks (other models may edge it there) or if you require a smaller-footprint model for mobile inference; production tool execution must still be implemented and sandboxed by the surrounding system.
Where it fits
Positioned between broad general-vision VLMs and specialized robotics perception stacks: it retains general visual competencies while layering explicit spatial/embodied reasoning and agent output capabilities useful for downstream planning and manipulation pipelines.
