**Cambrian-1: A Fully Open, Vision-Centric Exploration of Multimodal LLMs** This paper introduces Cambrian-1, a family of multimodal large language models (MLLMs) designed with a vision-centric approach. The authors address the gap between visual representation learning and the design of vision components in MLLMs, which often lacks exploration and integration. They evaluate various visual representations using MLLMs and visual instruction tuning, offering insights into different models and architectures. The study critically examines existing MLLM benchmarks and introduces a new vision-centric benchmark, CV-Bench, to better assess visual grounding in real-world scenarios. Key contributions include: 1. **Spatial Vision Aggregator (SVA)**: A dynamic and spatially-aware connector that integrates high-resolution vision features with LLMs while reducing the number of tokens. 2. **Instruction Tuning Data**: Curated high-quality visual instruction-tuning data from public sources, emphasizing the importance of data source balancing and distribution ratio. 3. **State-of-the-Art Performance**: Cambrian-1 achieves top performance across diverse benchmarks and excels in visual-centric tasks. The paper provides model weights, code, datasets, and detailed recipes for training and evaluation, aiming to inspire and accelerate advancements in multimodal systems and visual representation learning.**Cambrian-1: A Fully Open, Vision-Centric Exploration of Multimodal LLMs** This paper introduces Cambrian-1, a family of multimodal large language models (MLLMs) designed with a vision-centric approach. The authors address the gap between visual representation learning and the design of vision components in MLLMs, which often lacks exploration and integration. They evaluate various visual representations using MLLMs and visual instruction tuning, offering insights into different models and architectures. The study critically examines existing MLLM benchmarks and introduces a new vision-centric benchmark, CV-Bench, to better assess visual grounding in real-world scenarios. Key contributions include: 1. **Spatial Vision Aggregator (SVA)**: A dynamic and spatially-aware connector that integrates high-resolution vision features with LLMs while reducing the number of tokens. 2. **Instruction Tuning Data**: Curated high-quality visual instruction-tuning data from public sources, emphasizing the importance of data source balancing and distribution ratio. 3. **State-of-the-Art Performance**: Cambrian-1 achieves top performance across diverse benchmarks and excels in visual-centric tasks. The paper provides model weights, code, datasets, and detailed recipes for training and evaluation, aiming to inspire and accelerate advancements in multimodal systems and visual representation learning.