Meta today released Llama 3.2, a major update to its open-source model family that adds vision capabilities and introduces lightweight models designed to run on edge devices and mobile phones. The release includes two vision-language models at 11 billion and 90 billion parameters, as well as text-only models at 1 billion and 3 billion parameters that support a 128,000-token context window.

The 1B and 3B models, created through pruning and distillation from the larger Llama 3.1 8B model, are small enough to fit on-device while retaining strong performance on tasks like summarization, instruction following, and tool use. Meta says these models are enabled on day one for Qualcomm and MediaTek hardware and optimized for Arm processors. The vision models are designed as drop-in replacements for Llama 3.1 text models, adding image understanding capabilities without sacrificing existing language performance.

Alongside the models, Meta released the first official Llama Stack distributions — standardized APIs for inference, tool use, and retrieval-augmented generation — with support from partners including AWS, Databricks, Dell, Google Cloud, and NVIDIA. On-device distribution runs via PyTorch ExecuTorch, while single-node distribution is available through Ollama.

The announcement was immediately met with enthusiasm, with developer Simon Willison saying he is ‘absolutely amazed at how capable the new 1B model is’ given the 1.3GB download. The thread also saw skepticism about the vision models’ competitiveness, with some users noting that alternatives like Qwen-2-VL and Molmo appear to offer better performance on visual benchmarks. Meta acknowledged that the vision models were compared against Claude 3 Haiku and GPT-4o-mini.