Meta drops Llama 3.2 Vision, but EU users are blocked from the multimodal models

Meta just shipped Llama 3.2, and the headliner is a pair of vision models that give the open-source LLM a set of eyes. You get two sizes: an 11B parameter model tuned for consumer GPUs and a beastly 90B version for large-scale applications. Both come in base and instruction-tuned variants. Under the hood, they strap a vision tower and image adapter onto the existing Llama 3.1 text models—8B for the 11B vision model and 70B for the 90B—and Meta froze the text weights during training to avoid degrading language performance. The result is a model that can handle visual reasoning, document Q&A, and image-text retrieval with a 128k token context window. Chain-of-thought answers are a standout feature, making the visual reasoning feel less like a parlor trick and more like actual reasoning.

But here’s the kicker: if you’re an individual living in the European Union, or a company headquartered there, you can’t use the multimodal models. Meta’s acceptable use policy carves out the EU entirely. The restriction doesn’t trickle down to end users of products built with the models, so a startup in Paris can’t fine-tune Llama 3.2 Vision, but a company in Austin can ship a product to customers in Berlin. It’s a strange, territorial twist on “open” that’s going to frustrate a lot of European developers who’ve been all-in on the Llama ecosystem since the beginning. The text-only 1B and 3B models, designed for on-device use cases like summarization and tool use, don’t have this restriction.

On the smaller side of the release, those 1B and 3B text models are genuinely interesting for a different reason: they’re built to run locally. We’re talking on-device prompt rewriting, multilingual retrieval, and locally running assistants that don’t need to phone home. For tasks where latency and privacy matter more than raw intelligence, these tiny models outperform many open-access alternatives in their weight class. Meta also released a 1B version of Llama Guard, their safety classifier, so you can deploy a full stack—model plus guardrails—on edge hardware. The vision models integrate with Hugging Face Transformers and TGI out of the gate, and there’s a clear path to fine-tune the 11B Vision model on a single GPU using TRL.

The elephant in the room is what this EU exclusion means for the broader “open-weight” movement. Meta trained these vision models on a massive dataset of 6 billion image-text pairs, but they’re not saying much about what’s in that data. Combine that opacity with a geographic license restriction, and you’ve got something that feels less like open source and more like a strategic asset with a kill switch. It’s still a powerful tool, and the 11B model’s ability to run on a single consumer GPU will make it the default choice for a lot of builders. Just don’t call it open.