Meta AI Updates and Changelog

Meta held its first dedicated AI developer conference, separate from its broader Connect events. LlamaCon featured keynotes from Chief Product Officer Chris Cox and a conversation between Mark Zuckerberg and Microsoft CEO Satya Nadella. Announcements included the Llama API preview, fine-tuning tools for Llama 3.3 8B, CyberSecEval 4 security tools, the Llama Defenders Program for security evaluation, and $1.5 million in Llama Impact Grants to 10 international recipients. The event positioned Meta as a platform company for AI development, not just a model provider.

Meta launched a preview of the Llama API, allowing developers to build on Llama models directly through Meta's infrastructure. The API includes fine-tuning capabilities, evaluation tools, and a promise that models built on the platform are portable: developers can take them to any hosting provider. The security commitment was explicit: Meta stated it does not use prompts or model responses to train its AI models. Available initially to select customers with plans for broader rollout.

Meta released two Llama 4 models. Scout uses 17 billion active parameters with a 10 million token context window and 16 experts, targeting enterprise document analysis and long-context tasks. Maverick uses the same 17 billion active parameters but with 128 experts and a 1 million token context window, optimized for speed on coding, chatbots, and general-purpose tasks. Both use the Mixture-of-Experts architecture and support native multimodal input (text and images). A third model, Behemoth (288 billion active parameters across 2 trillion total), remained in training. The release marked Llama's transition from text-only to natively multimodal.

Meta expanded voice interaction with Meta AI across Messenger, Facebook, WhatsApp, and Instagram DMs. Users can speak to the assistant and receive spoken responses. The feature builds on the voice capabilities introduced with Llama 3.2 and represents Meta's bet that AI experiences will shift from text-first to voice-first as speech models become more natural.

Users can now share photos in chats with Meta AI, and the assistant understands visual content to answer questions about it. Combined with the Llama 4 multimodal architecture, this enables visual question answering, image description, and context-aware assistance based on what the user is looking at.

Meta AI on Ray-Ban Meta smart glasses rolled out in France, Italy, Ireland, and Spain. The custom Llama model running on the glasses lets users get information, identify objects, and complete tasks hands-free. The European expansion followed the US launch and positioned the glasses as a practical AI interface rather than a novelty device.

Meta expanded access to AI Studio, its platform for creating custom AI characters, to additional countries and languages including India, Pakistan, Mexico, Ecuador, Peru, Colombia, Argentina, and Chile. Hundreds of thousands of custom AIs have been created offering cooking tips, memes, affirmations, and more.

Meta released the Llama 3.3 70B instruction-tuned model as a text-only update with improved performance and reduced cost. The model targeted developers who needed a capable, efficient model for production deployment without multimodal requirements.

At Connect 2024, Meta introduced Llama 3.2 with multimodal features and models ranging from lightweight (1B and 3B parameters for edge devices) to capable (11B and 90B for vision tasks). The release included voice interaction for Meta AI, photo understanding in chat, and an experimental Reel translation feature. Business AIs on WhatsApp and Messenger also launched for click-to-message advertising.

The Llama 3 release included 8B and 70B parameter models alongside the launch of Meta AI as a standalone assistant across Facebook, Instagram, WhatsApp, Messenger, and the web. The tokenizer delivered up to 15% fewer tokens compared to Llama 2, improving efficiency. Meta positioned Meta AI as the world's leading AI assistant at launch, available in more countries across more surfaces than any competitor.

Multiple outlets reported that Meta was developing a closed-source AI model code-named Avocado, internally acknowledging that Llama 4's performance and adoption had not met expectations relative to competitors. The reports suggested Zuckerberg was considering a strategic pivot to compete with closed models from OpenAI, Google, and Anthropic using a closed model, while maintaining the open-source Llama line for community adoption. If confirmed, this would represent a significant shift in Meta's AI strategy.