One-RL-to-See-Them-All/Orsta-7B

VISIONPricing:Input $0.4 / Output $0.8Concurrent Unit Cost:1Model Size:7BQuant:FP8Context Size:32kPublished:May 25, 2025License:mitArchitecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Orsta-7B is a 7 billion parameter vision-language model (VLM) developed by MiniMax-AI, post-trained with the novel V-Triune unified reinforcement learning system. This model excels in both visual reasoning and visual perception tasks, having been jointly optimized across eight diverse visual tasks including mathematics, object detection, and OCR. Orsta-7B demonstrates superior performance on the MEGA-Bench Core benchmark, making it suitable for complex multimodal applications requiring robust visual understanding.

Loading preview...

Orsta-7B: A Vision-Language Model for Unified Visual Tasks

Orsta-7B is a 7 billion parameter vision-language model (VLM) developed by MiniMax-AI, distinguished by its training with the novel V-Triune unified reinforcement learning (RL) system. This system allows for joint optimization across diverse multimodal tasks within a single training pipeline, enhancing generalization and capabilities.

Key Capabilities

  • Unified Visual Task Optimization: Orsta-7B is trained on eight challenging visual tasks, encompassing both visual reasoning (Mathematics, Science QA, Chart Understanding, Puzzle Solving) and visual perception (Object Detection, Visual Grounding, OCR, Object Counting).
  • V-Triune RL Framework: Utilizes a Visual Triple-Unified Reinforcement Learning system with Sample-Level Data Formatting, Verifier-Level Reward Computation, and Source-Level Metric Monitoring, including an innovative Dynamic IoU reward for perception tasks. More details are available in the V-Triune paper.
  • Strong Performance: Achieves a MEGA-Bench Core score of 38.31, outperforming models like QwenVL-2.5-7B and InternVL-3-8B in its class, particularly excelling in Knowledge, Perception, and Information Extraction categories.

Good for

  • Applications requiring robust performance across a broad spectrum of visual reasoning and perception tasks.
  • Scenarios demanding precise object detection, visual grounding, and optical character recognition.
  • Complex multimodal understanding where intricate visual content needs to be related to textual prompts.

Finetunes
1 models