Meta releases Llama 3.2 with vision models and small models for phones

Meta/Llama

Meta released Llama 3.2: 11B and 90B vision models that reason about images and documents, and 1B and 3B text models with 128K-token context optimized to run on phones. The small models were optimized from day one for Qualcomm and MediaTek chips and Arm processors, enabling on-device agentic applications.

• Vision models: 11B and 90B for image reasoning, document analysis, competitive with Claude 3 Haiku and GPT-4o mini
• Lightweight models: 1B and 3B text; 128K token context; state-of-the-art on-device performance
• Optimized day one for Qualcomm, MediaTek, Arm; enables on-device privacy-preserving agentic applications
• Achieved through structured pruning and knowledge distillation from larger teacher models