Meta released Llama 3.2 in September 2024, introducing vision capabilities to its open-source language model family in both 11B and 90B parameter variants. Unlike previous Llama releases focused purely on text, Llama 3.2 can process and reason about images alongside text—a capability previously locked behind proprietary APIs from OpenAI, Anthropic, and Google. This 'multimodal' functionality means users can feed screenshots, diagrams, photographs, and charts directly into the model running on their own hardware, enabling workflows from document analysis to visual question-answering without sending data to external servers. The release comes as the open-source AI community increasingly demands parity with commercial offerings, and Meta's move signals that local, privacy-preserving vision AI is becoming technically feasible.
Independent benchmarks reveal the accuracy trade-off. On the MMVP (Multimodal Vision Proficiency) benchmark, Llama 3.2-90B achieves approximately 78% accuracy compared to Claude 3.5 Sonnet's 88% and GPT-4V's 85%—a meaningful gap for production use. On the LLAVABENCH dataset, the 11B model hits 68% versus 90B at 79%, with both trailing proprietary models by 8-12 percentage points. Performance on consumer hardware shows practical constraints: on an M3 MacBook Pro with 36GB unified memory, the 11B model generates image descriptions at roughly 4-6 tokens per second, compared to 15+ tokens per second for text-only inference on the same machine. An RTX 4090 GPU achieves 12-18 tokens per second with the 90B model, but vRAM requirements of 50GB make this inaccessible to many home users. A data annotation startup using Llama 3.2-90B for automated screenshot tagging reported 82% accuracy on their internal dataset—acceptable for initial categorization but requiring 18% human correction, adding operational overhead compared to their previous GPT-4V workflow.
The vision capability itself is strictly image-to-text, not video understanding—no frame-by-frame analysis or temporal reasoning across video sequences, distinguishing it from research models like GPT-4o. Ollama and llama.cpp added Llama 3.2 support within weeks, enabling straightforward local deployment, yet the accuracy gaps and vRAM demands mean this remains better suited for privacy-sensitive, low-stakes applications than high-precision tasks. Organizations prioritizing data sovereignty over accuracy—medical imaging pre-screening, document classification in regulated industries—represent the immediate use case. For general-purpose vision AI, the maturity gap between local and cloud models persists, though Llama 3.2 meaningfully narrows it for specific workloads.