Llama 3.2 Vision and on-device models released; EU license restrictions for multimodal variants

AI Impact Summary

Meta's Llama 3.2 release brings on-device multimodal capabilities (Llama 3.2 Vision) and on-device text models (1B/3B), plus Llama Guard 3 with vision support. The models are available as open-weight variants on Hugging Face Hub, with 128k context length for vision tasks and specific tile sizes (11B Vision uses 448 tiles; 90B Vision uses 560). Licensing changes block EU-domiciled entities from licensing multimodal Llama 3.2 models, which could constrain EU deployments even as non-EU customers can use them; plan for regional licensing compliance and distribution. Expect increased demand for offline/offline-enabled workflows and privacy-preserving AI, with guardrails via Llama Guard 3.