Meta Llama 3.2 Adds Vision Models and On-Device LLMs

Vision comes to Llama

Meta released Llama 3.2 on September 25, 2024 at its Connect event. The release added the first Llama models to support vision, the 11B and 90B, which handle image reasoning tasks like document-level understanding of charts and graphs, image captioning, and visual grounding such as pinpointing objects from a natural-language description. Meta said the vision models were competitive with Claude 3 Haiku and GPT-4o-mini on image understanding.

Lightweight on-device models

Llama 3.2 also introduced lightweight, text-only 1B and 3B models that fit onto edge and mobile devices, supporting 128K-token context and strong tool calling for on-device agentic apps where data never leaves the device. These models were enabled on day one for Qualcomm and MediaTek hardware and optimized for Arm processors, with Meta highlighting instant local responses and improved privacy.

How the models were built

For the vision models, Meta trained a set of adapter weights, a series of cross-attention layers, to feed a pretrained image encoder's representations into the pretrained Llama 3.1 language model without updating the language parameters, keeping text-only capabilities intact as a drop-in replacement. The 1B and 3B models were created from the Llama 3.1 8B using structured pruning and knowledge distillation from the 8B and 70B models.

Llama Stack and safety

Meta shared the first official Llama Stack distributions to simplify working with Llama across single-node, on-prem, cloud and on-device environments, with on-device distribution via PyTorch ExecuTorch and single-node via Ollama. New safeguards included Llama Guard 3 11B Vision for image-plus-text filtering and a pruned, quantized Llama Guard 3 1B shrunk from 2,858 MB to 438 MB for constrained deployments.

Related on Skillo

See also: Llama 3.1 and the 405B frontier model, PyTorch 2.4's Python 3.12 support.

Sources

• Llama 3.2: edge AI and vision (ai.meta.com)

Published date reflects the original event date (2024-09-25). This article is original Skillo editorial written from the sources above; facts were verified in September 2026.