Ollama now supports LLaVA 1.6 vision models in 7B, 13B, and a new 34B parameter size. The update brings support for up to 4x higher image resolution, improved text recognition and reasoning trained on document, chart, and diagram data, and more permissive licensing under Apache 2.0 or the LLaMA 2 Community License. Models can be run via CLI, Python, or JavaScript, with examples shown for object detection and text recognition from images.
Questions this post answers
How do I run the LLaVA 1.6 vision model locally with Ollama?
Use the ollama run command with the model tag, for example 'ollama run llava:7b', 'ollama run llava:13b', or 'ollama run llava:34b'. Pass an image with a prompt like 'ollama run llava "describe this image: ./art.jpg"' referencing a .jpg or .png file path directly in the command. daily.dev surfaces practical guides for running local multimodal models like LLaVA.
What improvements does LLaVA 1.6 add over previous LLaVA versions in Ollama?
LLaVA 1.6 supports up to 4x higher image resolution for capturing more detail, improved text recognition and reasoning from training on additional document, chart, and diagram datasets, and more permissive licensing distributed under Apache 2.0 or the LLaMA 2 Community License. It also adds a new 34B parameter size alongside the existing 7B and 13B models. Developers comparing vision model releases can track updates like this one on daily.dev.
How do I send an image to a vision model using the Ollama Python or JavaScript library?
Call ollama.chat with a messages array containing a user role, content text, and an images field listing the file path, such as {'role': 'user', 'content': 'Describe this image:', 'images': ['./art.jpg']}. Base64-encoded files are also accepted in the images parameter for the Python, JavaScript, and REST API interfaces. Those integrating vision models into apps can find implementation walkthroughs on daily.dev.