Ollama 0.4 Now Supports Building Llama 3.2 Vision + Visual RAG System

Ollama 0.4 now supports Llama 3.2 Vision for enhanced visual RAG retrieval. Discover its faster processing, improved OCR, and LocalGPT Vision integration.

In this article, I will introduce Ollama's recent update supporting Llama 3.2 Vision and share the performance results of Llama 3.2 Vision.

I will also introduce a visual RAG system, demonstrating how to integrate Llama 3.2 Vision with this system to complete tasks based on visual RAG retrieval.

Ollama 0.4.0 Update Highlights

• Support for the Llama 3.2 Vision (aka Mllama) architecture

• Faster subsequent requests to the visual model

• Fixed an issue where stop sequences were not correctly detected

• Ollama can now import models from Safetensors without needing a Modelfile (when running `ollama create my-model`)

• Fixed invalid character issues when redirecting output to files on Windows

• Resolved Ollama error due to invalid model data