After Meta said it would keep its next multimodal AI out of Europe, the company is now releasing Llama 3.1, an open-source AI model with performance rivaling ChatGPT and Google Gemini. The model now has 405 billion parameters, achieved with training on 16,000 enterprise-level Nvidia GPUs. What’s that mean for you, aside from some existential dread on how much money and energy that must have cost? Better chats in Meta AI and a new feature that can generate AI imagery starring you.
It’s the largest open-source AI model yet, and the company is already working with the likes of Amazon, Microsoft, and Google to help developers integrate it into their apps. It’s also free to download here, and developers can even use it to improve their own models. Non-developers are probably better off using the version already built into WhatsApp and Meta AI, although users will get bumped from the 405-billion parameter version to the 70-billion one if they bug the chatbot too much.
Credit: Meta
According to benchmark numbers released by Meta, Llama 3.1 performed within a few points of other models on most tasks, sometimes above and sometimes below, with GPT-4o and Claude 3.5 Sonnet offering the stiffest competition. Users are unlikely to notice a difference, but Meta is banking on the combination of performance and open-source availability wooing more developers to its side.
In addition to getting more parameters, Llama 3.1 also now supports French, German, Hindi, Italian, and Spanish. Instagram and Facebook are lagging a bit behind in getting the upgrade, but Meta says the social media sites will be updated with it in the coming weeks.
Right now, the flashiest way to test out Llama 3.1 is the new Imagine Me feature, which allows the AI to scan your face and then place it into AI imagery. You could ask the bot to turn you into an astronaut, or picture you with a different hair color. In a shot at Apple’s Genmoji, you can even use “Imagine Me” to generate stickers of yourself.
Meta is also making it possible to use Meta AI while making a Facebook or Instagram post, which could make it easier for the company to tag AI-generated images on the site, since posts made this way will already be marked. If you can’t beat ‘em, join ‘em, I suppose.
