Open-o3-Video-v2: Agentic Video Reasoning with Spatio-Temporal Evidence
October 12, 2026
Open-o3-Video-v2 treats video reasoning as an agentic process where the model identifies missing information and acquires additional observations. It uses the STGR-v2-CoT135K dataset to ground reasoning in specific timestamps and object-level bounding boxes.
HOW THIS AFFECTS YOU
●
builderYou can build video agents that provide verifiable evidence through timestamps and bounding boxes.
●
researcherThis moves video reasoning from sparse sampling toward a continuous, agentic evidence-grounding process.