Meta’s Llama 3.2 AI Can Run Locally and Understand Images

Meta just released the next generation of its free and open source LLMs. The new Llama 3.2 models can be run locally (even on mobile devices) and they’ve now gained image processing capabilities. Vision-supported multimodal LLMs that are also open source and free are still rare.

The Llama 3.2 comes in two medium-weight and two lightweight models. The 90B and 11BLLMscan process images like photos, documents, charts, and graphs. “For example, a person could ask a question about which month in the previous year their small business had the best sales, and Llama 3.2 can then reason based on an available graph and quickly provide the answer. In another example, the model could reason with a map and help answer questions such as when a hike might become steeper or the distance of a particular trail marked on the map,” Meta explained the new feature.

You can also ask these models to describe an image in detail or create stories or captions based on the description. The performance of these two models is comparable to theChatGPT 4o Miniand Claude 3 Haiku.

The lighter 1B and 3B Llama 3.2 models can run locally on a few edge devices (like smartphones and PCs). Meta has partnered with Qualcomm to run the Llama 3.2 models locally on Snapdragon 8 Gen 3 and Snapdragon X series chips. You’ll find these platforms on premium Android phones and the new generation ofCopilot+ PCs, respectively. These two models don’t have image recognition, but they can handle up to 128K tokens, allowing them to process long documents and keep conversations going for longer without losing context.

Running these models on the device is better for privacy and performance. Your data is never sent to the cloud, so an app using this model is more private by default. There’s minimal lag because the processing doesn’t involve the internet. Meta says the on-device responses “feel instantaneous.”

The Llama 3.2 models available for download onllama.comandHugging Face, and you can try them out on a desktop computer through LM Studio (just search for “Llama-3.2” in the Discover tab).