OraRL/Video-ORA-9B

VISIONConcurrent Unit Cost:1Model Size:9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 24, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Video-ORA-9B is a 9 billion parameter model developed by OraRL, built upon Qwen3.5-9B, designed for unified video understanding. It utilizes OraRL (Annotations as Rollouts) post-training to directly handle seven distinct video task families without requiring chain-of-thought decoding. With a native context length of 262,144 tokens, this model excels in tasks like temporal grounding, visual tracking, segmentation, and video question answering, providing task-native answers.

Loading preview...

Video-ORA-9B: Unified Video Understanding with OraRL

Video-ORA-9B is a 9 billion parameter model developed by OraRL, specifically engineered for unified video understanding. Built on the Qwen3.5-9B architecture, it leverages a novel OraRL (Annotations as Rollouts) post-training approach. This allows the model to directly process and provide task-native answers across a wide range of video-related tasks, eliminating the need for chain-of-thought decoding.

Key Capabilities

This model is designed to handle seven distinct task families within a single framework, offering direct solutions for:

  • Temporal grounding: Identifying specific events within a video's timeline.
  • Visual tracking: Following objects or subjects across video frames.
  • Image and video segmentation: Delineating objects and regions in visual media.
  • Spatial grounding: Locating objects or regions within a single frame.
  • Spatial-temporal grounding: Pinpointing objects across space and time.
  • Video question answering: Answering queries based on video content.
  • Spatial intelligence: Understanding spatial relationships and reasoning.

Differentiators and Performance

Video-ORA-9B stands out by providing direct, task-native answers without complex decoding strategies. It features a substantial native context length of 262,144 tokens, enabling comprehensive video analysis. Benchmarks indicate that Video-ORA-9B leads in matched seven-family comparisons against multimodal baselines, particularly when direct answers are required. The model is intended for research in structured image/video perception and benchmark evaluation, supporting direct answer prompts with enable_thinking=False for evaluation consistency. More details on benchmarks and protocols are available in the OraRL repository and paper.