Meta releases Llama 3.2 with vision and edge models

Meta released Llama 3.2 at its Connect developer conference, adding two capabilities the Llama line had lacked: vision and on-device operation. The 11B and 90B models were the first in the Llama family able to accept images, handling tasks such as reading charts and documents or grounding a description to a specific region of a picture. Separately, 1B and 3B text-only models were built for phones and other edge devices using pruning and distillation from larger Llama teacher models, and supported a 128,000-token context window.

Meta reported the vision models were competitive with Claude 3 Haiku and GPT-4o-mini on image-recognition benchmarks, and that the 3B model outperformed Google’s Gemma 2 2.6B and Microsoft’s Phi 3.5-mini on instruction-following and summarisation tasks. The models were made available for download on Meta’s own site and via Hugging Face under Meta’s Llama licence, which permits broad commercial use with restrictions above very large usage thresholds.

Llama 3.2 extended Meta’s open-weight strategy into two directions competitors were also pursuing that week — multimodal input, matched days apart by AI2’s Molmo, and small, efficient models for local deployment. Coming a week after Alibaba’s Qwen2.5 family and alongside Molmo, the release was one of several in a single September that broadened what “open-weight” models could do beyond text, intensifying the argument that openly available models could match closed, proprietary ones on core capabilities while running on hardware developers controlled themselves.