OpenMMReasoner/OpenMMReasoner-RL

VISIONPricing:Input $0.4 / Output $0.8Concurrent Unit Cost:1Model Size:7BQuant:FP8Context Size:32kPublished:Nov 19, 2025License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

OpenMMReasoner/OpenMMReasoner-RL is a 7 billion parameter multimodal reasoning model developed by Zhang, Kaichen et al. It utilizes a transparent two-stage training recipe involving supervised fine-tuning (SFT) on an 874K-sample dataset and reinforcement learning (RL) on a 74K-sample dataset. This model excels in multimodal reasoning tasks, achieving state-of-the-art results on benchmarks like MathVista, MathVerse, and WeMath, demonstrating robust performance with efficient data usage.

Loading preview...

Overview

OpenMMReasoner is a 7 billion parameter multimodal reasoning model that introduces a transparent, two-stage training recipe designed to enhance reasoning capabilities. This approach combines supervised fine-tuning (SFT) with reinforcement learning (RL) to achieve robust performance in multimodal domains.

Key Capabilities

  • Advanced Multimodal Reasoning: Excels at complex reasoning tasks involving both visual and textual information.
  • Efficient Training: Achieves strong results with significantly less data (874K SFT samples, 74K RL samples) compared to many competing methods, emphasizing data quality over sheer volume.
  • Benchmark Performance: Surpasses strong baselines, including a 11.6% improvement over Qwen2.5-VL-7B-Instruct across nine multimodal reasoning benchmarks. Notably, it achieves 79.5% on MathVista testmini, 63.8% on MathVerse testmini, and 79.0% on WeMath loose.
  • Transparent Methodology: Provides a fully transparent and reproducible data curation and training strategy, fostering scalable research.

Good for

  • Developing Multimodal AI Applications: Ideal for applications requiring sophisticated understanding and reasoning across different data types.
  • Research in Multimodal Reasoning: Serves as a strong empirical foundation for future research due to its transparent methodology and strong performance.
  • Benchmarking and Evaluation: Useful for evaluating and comparing multimodal reasoning models, given its competitive results on various benchmarks.

Finetunes
1 models