zlab-princeton/Vero-Qwen35-9B

VISIONConcurrent Unit Cost:1Model Size:9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:May 26, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Vero-Qwen35-9B is a 9 billion parameter open multimodal language model developed by zlab-princeton, part of the Vero family, specifically designed for general visual reasoning tasks. It excels across diverse categories including chart and OCR analysis, STEM problem-solving, spatial reasoning, knowledge recognition, grounding, counting, and instruction following. This model is fine-tuned using 600K curated RL samples and achieves state-of-the-art performance on the VeroEval 30-benchmark suite for visual reasoning.

Loading preview...

Vero-Qwen35-9B: An Open Multimodal Model for Visual Reasoning

Vero-Qwen35-9B, developed by zlab-princeton, is a 9 billion parameter open multimodal language model from the Vero family, focused on advancing general visual reasoning capabilities. It is part of a fully open release that includes models, training code, evaluation tools, and the Vero-600K dataset.

Key Capabilities & Features

  • Broad Visual Reasoning: Trained for extensive transfer across 6 visual reasoning categories, including chart and OCR interpretation, STEM problem-solving, spatial and action understanding, knowledge and recognition, grounding and counting, and captioning and instruction following.
  • Comprehensive Dataset: Utilizes Vero-600K, a dataset comprising 600,000 curated Reinforcement Learning (RL) samples derived from 59 distinct datasets.
  • State-of-the-Art Performance: Achieves state-of-the-art results among 8B parameter models on VeroEval, a robust 30-benchmark suite designed for general visual reasoning.
  • Base Model Improvement: Demonstrates improved performance across various base model families, including Qwen2.5-VL, Qwen3-VL, and MiMo-VL.

Ideal Use Cases

  • Complex Visual Analysis: Suited for applications requiring deep understanding and reasoning from visual inputs, such as interpreting charts, graphs, and scientific diagrams.
  • Multimodal Instruction Following: Effective for tasks where instructions involve both text and images, requiring the model to ground its responses in visual context.
  • Research and Development: Provides an open framework for researchers to explore and build upon advanced visual reasoning models, with access to models, data, and training code.